drivers/nvme/target/rdma.c | 331 +++++++++++++++++++++++++++++++++- drivers/nvme/target/tcp.c | 356 +++++++++++++++++++++++++++++++++++-- 2 files changed, 665 insertions(+), 22 deletions(-)
When running nvmf on SMP platform, current nvme target's RDMA and TCP use kworker to handle IO. But if there is other high workload in the system(eg: on kubernetes), the competition between the kworker and other workload is very radical. And since the kworker is scheduled by OS randomly, it's difficult to control OS resource and also tune the performance. If target support to use delicated polling task to handle IO, it's useful to control OS resource and gain good performance. So it makes sense to add polling task in rdma-rdma and rdma-tcp modules. Ping Gan (2): nvmet-rdma: add polling cq task for nvmet-rdma nvmet-tcp: add polling task for nvmet-tcp drivers/nvme/target/rdma.c | 331 +++++++++++++++++++++++++++++++++- drivers/nvme/target/tcp.c | 356 +++++++++++++++++++++++++++++++++++-- 2 files changed, 665 insertions(+), 22 deletions(-) -- 2.26.2
Hey Ping Gan, On 26/06/2024 11:28, Ping Gan wrote: > When running nvmf on SMP platform, current nvme target's RDMA and > TCP use kworker to handle IO. But if there is other high workload > in the system(eg: on kubernetes), the competition between the > kworker and other workload is very radical. And since the kworker > is scheduled by OS randomly, it's difficult to control OS resource > and also tune the performance. If target support to use delicated > polling task to handle IO, it's useful to control OS resource and > gain good performance. So it makes sense to add polling task in > rdma-rdma and rdma-tcp modules. This is NOT the way to go here. Both rdma and tcp are driven from workqueue context, which are bound workqueues. So there are two ways to go here: 1. Add generic port cpuset and use that to direct traffic to the appropriate set of cores (i.e. select an appropriate comp_vector for rdma and add an appropriate steering rule for tcp). 2. Add options to rdma/tcp to use UNBOUND workqueues, and allow users to control these UNBOUND workqueues cpumask via sysfs. (2) will not control interrupts to steer to other workloads cpus, but the handlers may run on a set of dedicated cpus. (1) is a better solution, but harder to implement. You also should look into nvmet-fc as well (and nvmet-loop for that matter).
>Hey Ping Gan, > > >On 26/06/2024 11:28, Ping Gan wrote: >> When running nvmf on SMP platform, current nvme target's RDMA and >> TCP use kworker to handle IO. But if there is other high workload >> in the system(eg: on kubernetes), the competition between the >> kworker and other workload is very radical. And since the kworker >> is scheduled by OS randomly, it's difficult to control OS resource >> and also tune the performance. If target support to use delicated >> polling task to handle IO, it's useful to control OS resource and >> gain good performance. So it makes sense to add polling task in >> rdma-rdma and rdma-tcp modules. > >This is NOT the way to go here. > >Both rdma and tcp are driven from workqueue context, which are bound >workqueues. > >So there are two ways to go here: >1. Add generic port cpuset and use that to direct traffic to the >appropriate set of cores >(i.e. select an appropriate comp_vector for rdma and add an appropriate >steering rule >for tcp). >2. Add options to rdma/tcp to use UNBOUND workqueues, and allow users >to >control >these UNBOUND workqueues cpumask via sysfs. > >(2) will not control interrupts to steer to other workloads cpus, but >the handlers may >run on a set of dedicated cpus. > >(1) is a better solution, but harder to implement. > >You also should look into nvmet-fc as well (and nvmet-loop for that >matter). hi Sagi Grimberg, Thanks for your reply, actually we had tried the first advice you suggested, but we found the performance was poor when using spdk as initiator. You know this patch is not only resolving OS resource competition issue, but also the perf issue. We have analyzed if we still use workqueue(kworker) as target when initiator is polling driver(eg: spdk), then workqueue/kworker target is the bottleneck since every nvmf request may have a wait latency from queuing on workqueue to begin processing, and the latency can be traced by wqlat of bcc (https://github.com/iovisor/bcc/blob/master/tools/wqlat.py). We think the latency is a disaster for the polling driver data plane, right? So we think adding a polling task mode on nvmet side to handle IO does really make sense; what's your opinion about this? And you mentioned we should also look into nvmet-fc, I agree with you. However currently we have no nvmf-fc's testbed; if we get the testbed, will do that. Thanks, Ping
On 01/07/2024 10:42, Ping Gan wrote: >> Hey Ping Gan, >> >> >> On 26/06/2024 11:28, Ping Gan wrote: >>> When running nvmf on SMP platform, current nvme target's RDMA and >>> TCP use kworker to handle IO. But if there is other high workload >>> in the system(eg: on kubernetes), the competition between the >>> kworker and other workload is very radical. And since the kworker >>> is scheduled by OS randomly, it's difficult to control OS resource >>> and also tune the performance. If target support to use delicated >>> polling task to handle IO, it's useful to control OS resource and >>> gain good performance. So it makes sense to add polling task in >>> rdma-rdma and rdma-tcp modules. >> This is NOT the way to go here. >> >> Both rdma and tcp are driven from workqueue context, which are bound >> workqueues. >> >> So there are two ways to go here: >> 1. Add generic port cpuset and use that to direct traffic to the >> appropriate set of cores >> (i.e. select an appropriate comp_vector for rdma and add an appropriate >> steering rule >> for tcp). >> 2. Add options to rdma/tcp to use UNBOUND workqueues, and allow users >> to >> control >> these UNBOUND workqueues cpumask via sysfs. >> >> (2) will not control interrupts to steer to other workloads cpus, but >> the handlers may >> run on a set of dedicated cpus. >> >> (1) is a better solution, but harder to implement. >> >> You also should look into nvmet-fc as well (and nvmet-loop for that >> matter). > hi Sagi Grimberg, > Thanks for your reply, actually we had tried the first advice you > suggested, but we found the performance was poor when using spdk > as initiator. I suggest that you focus on that instead of what you proposed. What is the source of your poor performance? > You know this patch is not only resolving OS resource > competition issue, but also the perf issue. We have analyzed if we > still use workqueue(kworker) as target when initiator is polling > driver(eg: spdk), then workqueue/kworker target is the bottleneck > since every nvmf request may have a wait latency from queuing on > workqueue to begin processing, That is incorrect, the work context polls the cq until it either drains it completely, or exhaust a quota of IB_POLL_BUDGET_WORKQUEUE (or NVMET_TCP_IO_WORK_BUDGET). Not every command gets its own workqueue queuing delay. And, what does the spdk initiator has to do with it? Didn't understand... > and the latency can be traced by wqlat > of bcc (https://github.com/iovisor/bcc/blob/master/tools/wqlat.py). > We think the latency is a disaster for the polling driver data plane, > right? If you need a target that polls all the time, you should probably resort to spdk. If there is room for optimization in nvmet we'll gladly take it, but this is not the way to go IMO. > So we think adding a polling task mode on nvmet side to handle > IO does really make sense; what's your opinion about this? I personally think that adding a polling kthread is questionable. However there is a precedent, io_uring sqthreads. So please look into what is done there. I don't mind having something like IB_POLL_IOTASK (or io_task threads in nvmet-tcp) if its done correctly (leverages common code). > And you > mentioned we should also look into nvmet-fc, I agree with you. > However currently we have no nvmf-fc's testbed; if we get the testbed, > will do that. There is fcloop, you should use that to test, same for loop. We don't want the transports to diverge in functionality.
>On 01/07/2024 10:42, Ping Gan wrote:
>>> Hey Ping Gan,
>>>
>>>
>>> On 26/06/2024 11:28, Ping Gan wrote:
>>>> When running nvmf on SMP platform, current nvme target's RDMA and
>>>> TCP use kworker to handle IO. But if there is other high workload
>>>> in the system(eg: on kubernetes), the competition between the
>>>> kworker and other workload is very radical. And since the kworker
>>>> is scheduled by OS randomly, it's difficult to control OS resource
>>>> and also tune the performance. If target support to use delicated
>>>> polling task to handle IO, it's useful to control OS resource and
>>>> gain good performance. So it makes sense to add polling task in
>>>> rdma-rdma and rdma-tcp modules.
>>> This is NOT the way to go here.
>>>
>>> Both rdma and tcp are driven from workqueue context, which are bound
>>> workqueues.
>>>
>>> So there are two ways to go here:
>>> 1. Add generic port cpuset and use that to direct traffic to the
>>> appropriate set of cores
>>> (i.e. select an appropriate comp_vector for rdma and add an
>>> appropriate
>>> steering rule
>>> for tcp).
>>> 2. Add options to rdma/tcp to use UNBOUND workqueues, and allow
>>> users
>>> to
>>> control
>>> these UNBOUND workqueues cpumask via sysfs.
>>>
>>> (2) will not control interrupts to steer to other workloads cpus,
>>> but
>>> the handlers may
>>> run on a set of dedicated cpus.
>>>
>>> (1) is a better solution, but harder to implement.
>>>
>>> You also should look into nvmet-fc as well (and nvmet-loop for that
>>> matter).
>> hi Sagi Grimberg,
>> Thanks for your reply, actually we had tried the first advice you
>> suggested, but we found the performance was poor when using spdk
>> as initiator.
>
>I suggest that you focus on that instead of what you proposed.
>What is the source of your poor performance?
Before these patches, we had used linux's RPS to forward the packets
to a fixed cpu set for nvmet-tcp. But when did that we can still not
cancel the competition between softirq and workqueue since nvme target's
kworker cpu core bind on socket's cpu which is from skb. Besides that
we found workqueue's wait latency was very high even we enabled polling
on nvmet-tcp by module parameter idle_poll_period_usecs. So when
initiator
is polling mode, the target of workqueue is the bottleneck. Below is
work's wait latency trace log of our test on our cluster(per node uses
4 numas 96 cores, 192G memory, one dual ports mellanox CX4LX(25Gbps X 2)
ethernet adapter and randrw 1M IO size) by RPS to 6 cpu cores. And
system's CPU and memory were used about 80%.
ogden-brown:~ #/usr/share/bcc/tools/wqlat -T -w nvmet_tcp_wq 1 2
01:06:59
usecs : count distribution
0 -> 1 : 0 | |
2 -> 3 : 0 | |
4 -> 7 : 0 | |
8 -> 15 : 3 | |
16 -> 31 : 10 | |
32 -> 63 : 3 | |
64 -> 127 : 2 | |
128 -> 255 : 0 | |
256 -> 511 : 5 | |
512 -> 1023 : 12 | |
1024 -> 2047 : 26 |* |
2048 -> 4095 : 34 |* |
4096 -> 8191 : 350 |************ |
8192 -> 16383 : 625 |******************************|
16384 -> 32767 : 244 |********* |
32768 -> 65535 : 39 |* |
01:07:00
usecs : count distribution
0 -> 1 : 1 | |
2 -> 3 : 0 | |
4 -> 7 : 4 | |
8 -> 15 : 3 | |
16 -> 31 : 8 | |
32 -> 63 : 10 | |
64 -> 127 : 3 | |
128 -> 255 : 6 | |
256 -> 511 : 8 | |
512 -> 1023 : 20 |* |
1024 -> 2047 : 19 |* |
2048 -> 4095 : 57 |** |
4096 -> 8191 : 325 |**************** |
8192 -> 16383 : 647 |******************************|
16384 -> 32767 : 228 |*********** |
32768 -> 65535 : 43 |** |
65536 -> 131071 : 1 | |
And the bandwidth of a node is only 3100MB. While we used the patch
and enable 6 polling task, the bandwidth can be 4000MB. It's a good
improvement.
>> You know this patch is not only resolving OS resource
>> competition issue, but also the perf issue. We have analyzed if we
>> still use workqueue(kworker) as target when initiator is polling
>> driver(eg: spdk), then workqueue/kworker target is the bottleneck
>> since every nvmf request may have a wait latency from queuing on
>> workqueue to begin processing,
>
>That is incorrect, the work context polls the cq until it either drains
>it
>completely, or exhaust a quota of IB_POLL_BUDGET_WORKQUEUE (or
>NVMET_TCP_IO_WORK_BUDGET). Not every command gets its own workqueue
>queuing delay.
>
>And, what does the spdk initiator has to do with it? Didn't
>understand...
Yes, target workqueue implementation will poll a quota; but when the
work
load was high we found many work will wait too long(some of them at
several
ms to hundred ms shown above histogram). We use the spdk initiator(by
polling mode) to send IO's read/write to nvme disks of a kubernetes
cluster's remote node.
>> and the latency can be traced by wqlat
>> of bcc (https://github.com/iovisor/bcc/blob/master/tools/wqlat.py).
>> We think the latency is a disaster for the polling driver data plane,
>> right?
>
>If you need a target that polls all the time, you should probably
>resort
>to spdk.
>If there is room for optimization in nvmet we'll gladly take it, but
>this is not the
>way to go IMO.
Yes, in the begining we did use the spdk as polling target driver,
but we suffered from spdk target could not support disk hot plug/unplug
well, sometimes it will cause data loss when did disk hot plug/unplug.
So we switch to kernel target driver because in production customer's
data security is first priority. And for kernel's target it has no
polling mode target driver, so we implemented these patches.
>> So we think adding a polling task mode on nvmet side to handle
>> IO does really make sense; what's your opinion about this?
>
>I personally think that adding a polling kthread is questionable.
>However there is a precedent, io_uring sqthreads. So please look
>into what is done there. I don't mind having something like
>IB_POLL_IOTASK (or
>io_task threads in nvmet-tcp) if its done correctly (leverages common
>code).
Yes, we have studied io_uring's code before implementing the patches.
Actually we followed io_uring's design idea in these patches.
>> And you
>> mentioned we should also look into nvmet-fc, I agree with you.
>> However currently we have no nvmf-fc's testbed; if we get the
>> testbed,
>> will do that.
>
>There is fcloop, you should use that to test, same for loop. We don't
>want
>the transports to diverge in functionality.
Ok, I will try, would you please give me some configuration guide for
fcloop since I never used fcloop before.
Thanks,
Ping
On 02/07/2024 13:02, Ping Gan wrote: >> On 01/07/2024 10:42, Ping Gan wrote: >>>> Hey Ping Gan, >>>> >>>> >>>> On 26/06/2024 11:28, Ping Gan wrote: >>>>> When running nvmf on SMP platform, current nvme target's RDMA and >>>>> TCP use kworker to handle IO. But if there is other high workload >>>>> in the system(eg: on kubernetes), the competition between the >>>>> kworker and other workload is very radical. And since the kworker >>>>> is scheduled by OS randomly, it's difficult to control OS resource >>>>> and also tune the performance. If target support to use delicated >>>>> polling task to handle IO, it's useful to control OS resource and >>>>> gain good performance. So it makes sense to add polling task in >>>>> rdma-rdma and rdma-tcp modules. >>>> This is NOT the way to go here. >>>> >>>> Both rdma and tcp are driven from workqueue context, which are bound >>>> workqueues. >>>> >>>> So there are two ways to go here: >>>> 1. Add generic port cpuset and use that to direct traffic to the >>>> appropriate set of cores >>>> (i.e. select an appropriate comp_vector for rdma and add an >>>> appropriate >>>> steering rule >>>> for tcp). >>>> 2. Add options to rdma/tcp to use UNBOUND workqueues, and allow >>>> users >>>> to >>>> control >>>> these UNBOUND workqueues cpumask via sysfs. >>>> >>>> (2) will not control interrupts to steer to other workloads cpus, >>>> but >>>> the handlers may >>>> run on a set of dedicated cpus. >>>> >>>> (1) is a better solution, but harder to implement. >>>> >>>> You also should look into nvmet-fc as well (and nvmet-loop for that >>>> matter). >>> hi Sagi Grimberg, >>> Thanks for your reply, actually we had tried the first advice you >>> suggested, but we found the performance was poor when using spdk >>> as initiator. >> I suggest that you focus on that instead of what you proposed. >> What is the source of your poor performance? > Before these patches, we had used linux's RPS to forward the packets > to a fixed cpu set for nvmet-tcp. But when did that we can still not > cancel the competition between softirq and workqueue since nvme target's > kworker cpu core bind on socket's cpu which is from skb. Besides that > we found workqueue's wait latency was very high even we enabled polling > on nvmet-tcp by module parameter idle_poll_period_usecs. So when > initiator > is polling mode, the target of workqueue is the bottleneck. Below is > work's wait latency trace log of our test on our cluster(per node uses > 4 numas 96 cores, 192G memory, one dual ports mellanox CX4LX(25Gbps X 2) > ethernet adapter and randrw 1M IO size) by RPS to 6 cpu cores. And > system's CPU and memory were used about 80%. I'd try a simple unbound CPU case, steer packets to say cores [0-5] and assign the cpumask of the unbound workqueue to cores [6-11]. > ogden-brown:~ #/usr/share/bcc/tools/wqlat -T -w nvmet_tcp_wq 1 2 > 01:06:59 > usecs : count distribution > 0 -> 1 : 0 | | > 2 -> 3 : 0 | | > 4 -> 7 : 0 | | > 8 -> 15 : 3 | | > 16 -> 31 : 10 | | > 32 -> 63 : 3 | | > 64 -> 127 : 2 | | > 128 -> 255 : 0 | | > 256 -> 511 : 5 | | > 512 -> 1023 : 12 | | > 1024 -> 2047 : 26 |* | > 2048 -> 4095 : 34 |* | > 4096 -> 8191 : 350 |************ | > 8192 -> 16383 : 625 |******************************| > 16384 -> 32767 : 244 |********* | > 32768 -> 65535 : 39 |* | > > 01:07:00 > usecs : count distribution > 0 -> 1 : 1 | | > 2 -> 3 : 0 | | > 4 -> 7 : 4 | | > 8 -> 15 : 3 | | > 16 -> 31 : 8 | | > 32 -> 63 : 10 | | > 64 -> 127 : 3 | | > 128 -> 255 : 6 | | > 256 -> 511 : 8 | | > 512 -> 1023 : 20 |* | > 1024 -> 2047 : 19 |* | > 2048 -> 4095 : 57 |** | > 4096 -> 8191 : 325 |**************** | > 8192 -> 16383 : 647 |******************************| > 16384 -> 32767 : 228 |*********** | > 32768 -> 65535 : 43 |** | > 65536 -> 131071 : 1 | | > > And the bandwidth of a node is only 3100MB. While we used the patch > and enable 6 polling task, the bandwidth can be 4000MB. It's a good > improvement. I think you will see similar performance with unbound workqueue and rps. > >>> You know this patch is not only resolving OS resource >>> competition issue, but also the perf issue. We have analyzed if we >>> still use workqueue(kworker) as target when initiator is polling >>> driver(eg: spdk), then workqueue/kworker target is the bottleneck >>> since every nvmf request may have a wait latency from queuing on >>> workqueue to begin processing, >> That is incorrect, the work context polls the cq until it either drains >> it >> completely, or exhaust a quota of IB_POLL_BUDGET_WORKQUEUE (or >> NVMET_TCP_IO_WORK_BUDGET). Not every command gets its own workqueue >> queuing delay. >> >> And, what does the spdk initiator has to do with it? Didn't >> understand... > Yes, target workqueue implementation will poll a quota; but when the > work > load was high we found many work will wait too long(some of them at > several > ms to hundred ms shown above histogram). We use the spdk initiator(by > polling mode) to send IO's read/write to nvme disks of a kubernetes > cluster's remote node. The initiator is an orthogonal detail here. the same issue exists regardless of spdk afaiu. Let's ignore it, its confusing. > >>> and the latency can be traced by wqlat >>> of bcc (https://github.com/iovisor/bcc/blob/master/tools/wqlat.py). >>> We think the latency is a disaster for the polling driver data plane, >>> right? >> If you need a target that polls all the time, you should probably >> resort >> to spdk. >> If there is room for optimization in nvmet we'll gladly take it, but >> this is not the >> way to go IMO. > Yes, in the begining we did use the spdk as polling target driver, > but we suffered from spdk target could not support disk hot plug/unplug > well, sometimes it will cause data loss when did disk hot plug/unplug. > So we switch to kernel target driver because in production customer's > data security is first priority. And for kernel's target it has no > polling mode target driver, so we implemented these patches. Well, its a hard sell for upstream nvmet. > >>> So we think adding a polling task mode on nvmet side to handle >>> IO does really make sense; what's your opinion about this? >> I personally think that adding a polling kthread is questionable. >> However there is a precedent, io_uring sqthreads. So please look >> into what is done there. I don't mind having something like >> IB_POLL_IOTASK (or >> io_task threads in nvmet-tcp) if its done correctly (leverages common >> code). > Yes, we have studied io_uring's code before implementing the patches. > Actually we followed io_uring's design idea in these patches. I'm talking about reusing what io_uring sqpoll tasks. Move them to common code, generalizing it to address what you need, and reuse that. Implementing a half-baked inspired version in nvmet is not going to fly here. Sorry.
> On 02/07/2024 13:02, Ping Gan wrote: >>> On 01/07/2024 10:42, Ping Gan wrote: >>>>> Hey Ping Gan, >>>>> >>>>> >>>>> On 26/06/2024 11:28, Ping Gan wrote: >>>>>> When running nvmf on SMP platform, current nvme target's RDMA and >>>>>> TCP use kworker to handle IO. But if there is other high workload >>>>>> in the system(eg: on kubernetes), the competition between the >>>>>> kworker and other workload is very radical. And since the kworker >>>>>> is scheduled by OS randomly, it's difficult to control OS >>>>>> resource >>>>>> and also tune the performance. If target support to use delicated >>>>>> polling task to handle IO, it's useful to control OS resource and >>>>>> gain good performance. So it makes sense to add polling task in >>>>>> rdma-rdma and rdma-tcp modules. >>>>> This is NOT the way to go here. >>>>> >>>>> Both rdma and tcp are driven from workqueue context, which are >>>>> bound >>>>> workqueues. >>>>> >>>>> So there are two ways to go here: >>>>> 1. Add generic port cpuset and use that to direct traffic to the >>>>> appropriate set of cores >>>>> (i.e. select an appropriate comp_vector for rdma and add an >>>>> appropriate >>>>> steering rule >>>>> for tcp). >>>>> 2. Add options to rdma/tcp to use UNBOUND workqueues, and allow >>>>> users >>>>> to >>>>> control >>>>> these UNBOUND workqueues cpumask via sysfs. >>>>> >>>>> (2) will not control interrupts to steer to other workloads cpus, >>>>> but >>>>> the handlers may >>>>> run on a set of dedicated cpus. >>>>> >>>>> (1) is a better solution, but harder to implement. >>>>> >>>>> You also should look into nvmet-fc as well (and nvmet-loop for >>>>> that >>>>> matter). >>>> hi Sagi Grimberg, >>>> Thanks for your reply, actually we had tried the first advice you >>>> suggested, but we found the performance was poor when using spdk >>>> as initiator. >>> I suggest that you focus on that instead of what you proposed. >>> What is the source of your poor performance? >> Before these patches, we had used linux's RPS to forward the packets >> to a fixed cpu set for nvmet-tcp. But when did that we can still not >> cancel the competition between softirq and workqueue since nvme >> target's >> kworker cpu core bind on socket's cpu which is from skb. Besides that >> we found workqueue's wait latency was very high even we enabled >> polling >> on nvmet-tcp by module parameter idle_poll_period_usecs. So when >> initiator >> is polling mode, the target of workqueue is the bottleneck. Below is >> work's wait latency trace log of our test on our cluster(per node >> uses >> 4 numas 96 cores, 192G memory, one dual ports mellanox CX4LX(25Gbps X >> 2) >> ethernet adapter and randrw 1M IO size) by RPS to 6 cpu cores. And >> system's CPU and memory were used about 80%. > I'd try a simple unbound CPU case, steer packets to say cores [0-5] > and > assign > the cpumask of the unbound workqueue to cores [6-11]. Okay, thanks for your guide. >> ogden-brown:~ #/usr/share/bcc/tools/wqlat -T -w nvmet_tcp_wq 1 2 >> 01:06:59 >> usecs : count distribution >> 0 -> 1 : 0 | | >> 2 -> 3 : 0 | | >> 4 -> 7 : 0 | | >> 8 -> 15 : 3 | | >> 16 -> 31 : 10 | | >> 32 -> 63 : 3 | | >> 64 -> 127 : 2 | | >> 128 -> 255 : 0 | | >> 256 -> 511 : 5 | | >> 512 -> 1023 : 12 | | >> 1024 -> 2047 : 26 |* | >> 2048 -> 4095 : 34 |* | >> 4096 -> 8191 : 350 |************ | >> 8192 -> 16383 : 625 |******************************| >> 16384 -> 32767 : 244 |********* | >> 32768 -> 65535 : 39 |* | >> >> 01:07:00 >> usecs : count distribution >> 0 -> 1 : 1 | | >> 2 -> 3 : 0 | | >> 4 -> 7 : 4 | | >> 8 -> 15 : 3 | | >> 16 -> 31 : 8 | | >> 32 -> 63 : 10 | | >> 64 -> 127 : 3 | | >> 128 -> 255 : 6 | | >> 256 -> 511 : 8 | | >> 512 -> 1023 : 20 |* | >> 1024 -> 2047 : 19 |* | >> 2048 -> 4095 : 57 |** | >> 4096 -> 8191 : 325 |**************** | >> 8192 -> 16383 : 647 |******************************| >> 16384 -> 32767 : 228 |*********** | >> 32768 -> 65535 : 43 |** | >> 65536 -> 131071 : 1 | | >> >> And the bandwidth of a node is only 3100MB. While we used the patch >> and enable 6 polling task, the bandwidth can be 4000MB. It's a good >> improvement. > > I think you will see similar performance with unbound workqueue and > rps. Yes, I remodified the nvmet-tcp/nvmet-rdma code for supporting unbound workqueue, and in same prerequisites of above to run test, and compared the result of unbound workqueue and polling mode task. And I got a good performance for unbound workqueue. For unbound workqueue TCP we got 3850M/node, it's almost equal to polling task. And also tested nvmet-rdma we get 5100M/node for unbound workqueue RDMA versus 5600M for polling task, seems the diff is very small. Anyway, your advice is good. Do you think we should submit the unbound workqueue patches for nvmet-tcp and nvmet-rdma to upstream nvmet? BTW I have another question: Will nvmet of upstream have the plan to support polling queue when doing submit_bio in future? >> >>>> You know this patch is not only resolving OS resource >>>> competition issue, but also the perf issue. We have analyzed if we >>>> still use workqueue(kworker) as target when initiator is polling >>>> driver(eg: spdk), then workqueue/kworker target is the bottleneck >>>> since every nvmf request may have a wait latency from queuing on >>>> workqueue to begin processing, >>> That is incorrect, the work context polls the cq until it either >>> drains >>> it >>> completely, or exhaust a quota of IB_POLL_BUDGET_WORKQUEUE (or >>> NVMET_TCP_IO_WORK_BUDGET). Not every command gets its own workqueue >>> queuing delay. >>> >>> And, what does the spdk initiator has to do with it? Didn't >>> understand... >> Yes, target workqueue implementation will poll a quota; but when the >> work >> load was high we found many work will wait too long(some of them at >> several >> ms to hundred ms shown above histogram). We use the spdk initiator(by >> polling mode) to send IO's read/write to nvme disks of a kubernetes >> cluster's remote node. > > The initiator is an orthogonal detail here. the same issue exists > regardless of > spdk afaiu. Let's ignore it, its confusing. > >> >>>> and the latency can be traced by wqlat >>>> of bcc (https://github.com/iovisor/bcc/blob/master/tools/wqlat.py). >>>> We think the latency is a disaster for the polling driver data >>>> plane, >>>> right? >>> If you need a target that polls all the time, you should probably >>> resort >>> to spdk. >>> If there is room for optimization in nvmet we'll gladly take it, but >>> this is not the >>> way to go IMO. >> Yes, in the begining we did use the spdk as polling target driver, >> but we suffered from spdk target could not support disk hot >> plug/unplug >> well, sometimes it will cause data loss when did disk hot >> plug/unplug. >> So we switch to kernel target driver because in production customer's >> data security is first priority. And for kernel's target it has no >> polling mode target driver, so we implemented these patches. > > Well, its a hard sell for upstream nvmet. >> >> >>>> So we think adding a polling task mode on nvmet side to handle >>>> IO does really make sense; what's your opinion about this? >>> I personally think that adding a polling kthread is questionable. >>> However there is a precedent, io_uring sqthreads. So please look >>> into what is done there. I don't mind having something like >>> IB_POLL_IOTASK (or >>> io_task threads in nvmet-tcp) if its done correctly (leverages >>> common >>> code). >> Yes, we have studied io_uring's code before implementing the patches. >> Actually we followed io_uring's design idea in these patches. > > I'm talking about reusing what io_uring sqpoll tasks. Move them to > common > code, generalizing it to address what you need, and reuse that. > Implementing a > half-baked inspired version in nvmet is not going to fly here. Sorry. Okay, got it. Thanks, Ping
On 7/4/24 10:10, Ping Gan wrote: >> On 02/07/2024 13:02, Ping Gan wrote: [ .. ] >>> And the bandwidth of a node is only 3100MB. While we used the patch >>> and enable 6 polling task, the bandwidth can be 4000MB. It's a good >>> improvement. >> >> I think you will see similar performance with unbound workqueue and >> rps. > > Yes, I remodified the nvmet-tcp/nvmet-rdma code for supporting unbound > workqueue, and in same prerequisites of above to run test, and compared > the result of unbound workqueue and polling mode task. And I got a good > performance for unbound workqueue. For unbound workqueue TCP we got > 3850M/node, it's almost equal to polling task. And also tested > nvmet-rdma we get 5100M/node for unbound workqueue RDMA versus 5600M for > polling task, seems the diff is very small. Anyway, your advice is good. > Do you think we should submit the unbound workqueue patches for nvmet-tcp > and nvmet-rdma to upstream nvmet? Please do. I have been using pretty much the same patch during development of my nvme-tcp scalability patchset, and using WQ_UNBOUND definitely improves the situation here. Cheers, Hannes -- Dr. Hannes Reinecke Kernel Storage Architect hare@suse.de +49 911 74053 688 SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
> On 7/4/24 10:10, Ping Gan wrote: >>> On 02/07/2024 13:02, Ping Gan wrote: > >>>> And the bandwidth of a node is only 3100MB. While we used the patch >>>> and enable 6 polling task, the bandwidth can be 4000MB. It's a good >>>> improvement. >>> >>> I think you will see similar performance with unbound workqueue and >>> rps. >> >> Yes, I remodified the nvmet-tcp/nvmet-rdma code for supporting >> unbound >> workqueue, and in same prerequisites of above to run test, and >> compared >> the result of unbound workqueue and polling mode task. And I got a >> good >> performance for unbound workqueue. For unbound workqueue TCP we got >> 3850M/node, it's almost equal to polling task. And also tested >> nvmet-rdma we get 5100M/node for unbound workqueue RDMA versus 5600M >> for >> polling task, seems the diff is very small. Anyway, your advice is >> good. >> Do you think we should submit the unbound workqueue patches for >> nvmet-tcp >> and nvmet-rdma to upstream nvmet? > > Please do. I have been using pretty much the same patch during > development of my nvme-tcp scalability patchset, and using WQ_UNBOUND > definitely improves the situation here. Thanks for your confirm! Okay, will do that. Regards, Ping
On 7/4/24 11:10, Ping Gan wrote: >> On 02/07/2024 13:02, Ping Gan wrote: >>>> On 01/07/2024 10:42, Ping Gan wrote: >>>>>> Hey Ping Gan, >>>>>> >>>>>> >>>>>> On 26/06/2024 11:28, Ping Gan wrote: >>>>>>> When running nvmf on SMP platform, current nvme target's RDMA and >>>>>>> TCP use kworker to handle IO. But if there is other high workload >>>>>>> in the system(eg: on kubernetes), the competition between the >>>>>>> kworker and other workload is very radical. And since the kworker >>>>>>> is scheduled by OS randomly, it's difficult to control OS >>>>>>> resource >>>>>>> and also tune the performance. If target support to use delicated >>>>>>> polling task to handle IO, it's useful to control OS resource and >>>>>>> gain good performance. So it makes sense to add polling task in >>>>>>> rdma-rdma and rdma-tcp modules. >>>>>> This is NOT the way to go here. >>>>>> >>>>>> Both rdma and tcp are driven from workqueue context, which are >>>>>> bound >>>>>> workqueues. >>>>>> >>>>>> So there are two ways to go here: >>>>>> 1. Add generic port cpuset and use that to direct traffic to the >>>>>> appropriate set of cores >>>>>> (i.e. select an appropriate comp_vector for rdma and add an >>>>>> appropriate >>>>>> steering rule >>>>>> for tcp). >>>>>> 2. Add options to rdma/tcp to use UNBOUND workqueues, and allow >>>>>> users >>>>>> to >>>>>> control >>>>>> these UNBOUND workqueues cpumask via sysfs. >>>>>> >>>>>> (2) will not control interrupts to steer to other workloads cpus, >>>>>> but >>>>>> the handlers may >>>>>> run on a set of dedicated cpus. >>>>>> >>>>>> (1) is a better solution, but harder to implement. >>>>>> >>>>>> You also should look into nvmet-fc as well (and nvmet-loop for >>>>>> that >>>>>> matter). >>>>> hi Sagi Grimberg, >>>>> Thanks for your reply, actually we had tried the first advice you >>>>> suggested, but we found the performance was poor when using spdk >>>>> as initiator. >>>> I suggest that you focus on that instead of what you proposed. >>>> What is the source of your poor performance? >>> Before these patches, we had used linux's RPS to forward the packets >>> to a fixed cpu set for nvmet-tcp. But when did that we can still not >>> cancel the competition between softirq and workqueue since nvme >>> target's >>> kworker cpu core bind on socket's cpu which is from skb. Besides that >>> we found workqueue's wait latency was very high even we enabled >>> polling >>> on nvmet-tcp by module parameter idle_poll_period_usecs. So when >>> initiator >>> is polling mode, the target of workqueue is the bottleneck. Below is >>> work's wait latency trace log of our test on our cluster(per node >>> uses >>> 4 numas 96 cores, 192G memory, one dual ports mellanox CX4LX(25Gbps X >>> 2) >>> ethernet adapter and randrw 1M IO size) by RPS to 6 cpu cores. And >>> system's CPU and memory were used about 80%. >> I'd try a simple unbound CPU case, steer packets to say cores [0-5] >> and >> assign >> the cpumask of the unbound workqueue to cores [6-11]. > Okay, thanks for your guide. > >>> ogden-brown:~ #/usr/share/bcc/tools/wqlat -T -w nvmet_tcp_wq 1 2 >>> 01:06:59 >>> usecs : count distribution >>> 0 -> 1 : 0 | | >>> 2 -> 3 : 0 | | >>> 4 -> 7 : 0 | | >>> 8 -> 15 : 3 | | >>> 16 -> 31 : 10 | | >>> 32 -> 63 : 3 | | >>> 64 -> 127 : 2 | | >>> 128 -> 255 : 0 | | >>> 256 -> 511 : 5 | | >>> 512 -> 1023 : 12 | | >>> 1024 -> 2047 : 26 |* | >>> 2048 -> 4095 : 34 |* | >>> 4096 -> 8191 : 350 |************ | >>> 8192 -> 16383 : 625 |******************************| >>> 16384 -> 32767 : 244 |********* | >>> 32768 -> 65535 : 39 |* | >>> >>> 01:07:00 >>> usecs : count distribution >>> 0 -> 1 : 1 | | >>> 2 -> 3 : 0 | | >>> 4 -> 7 : 4 | | >>> 8 -> 15 : 3 | | >>> 16 -> 31 : 8 | | >>> 32 -> 63 : 10 | | >>> 64 -> 127 : 3 | | >>> 128 -> 255 : 6 | | >>> 256 -> 511 : 8 | | >>> 512 -> 1023 : 20 |* | >>> 1024 -> 2047 : 19 |* | >>> 2048 -> 4095 : 57 |** | >>> 4096 -> 8191 : 325 |**************** | >>> 8192 -> 16383 : 647 |******************************| >>> 16384 -> 32767 : 228 |*********** | >>> 32768 -> 65535 : 43 |** | >>> 65536 -> 131071 : 1 | | >>> >>> And the bandwidth of a node is only 3100MB. While we used the patch >>> and enable 6 polling task, the bandwidth can be 4000MB. It's a good >>> improvement. >> I think you will see similar performance with unbound workqueue and >> rps. > Yes, I remodified the nvmet-tcp/nvmet-rdma code for supporting unbound > workqueue, and in same prerequisites of above to run test, and compared > the result of unbound workqueue and polling mode task. And I got a good > performance for unbound workqueue. For unbound workqueue TCP we got > 3850M/node, it's almost equal to polling task. And also tested > nvmet-rdma > we get 5100M/node for unbound workqueue RDMA versus 5600M for polling > task, > seems the diff is very small. Anyway, your advice is good. I'm a bit surprised that you see ~10% delta here. I would look into what is the root-cause of this difference. If indeed the load is high, the overhead of the workqueue mgmt should be negligible. I'm assuming you used IB_POLL_UNBOUND_WORKQUEUE ? > Do you think > we > should submit the unbound workqueue patches for nvmet-tcp and nvmet-rdma > to upstream nvmet? For nvmet-tcp, I think there is merit to split socket processing from napi context. For nvmet-rdma I think the only difference is if you have multiple CQs assigned with the same comp_vector. How many queues do you have in your test? > BTW I have another question: Will nvmet of upstream have the plan to > support > polling queue when doing submit_bio in future? No plans that I know of. Don't have a coherent idea of how that would work.
> On 7/4/24 11:10, Ping Gan wrote:
>>> On 02/07/2024 13:02, Ping Gan wrote:
>>>>> On 01/07/2024 10:42, Ping Gan wrote:
>>>>>>> Hey Ping Gan,
>>>>>>>
>>>>>>>
>>>>>>> On 26/06/2024 11:28, Ping Gan wrote:
>>>>>>>> When running nvmf on SMP platform, current nvme target's RDMA
>>>>>>>> and
>>>>>>>> TCP use kworker to handle IO. But if there is other high
>>>>>>>> workload
>>>>>>>> in the system(eg: on kubernetes), the competition between the
>>>>>>>> kworker and other workload is very radical. And since the
>>>>>>>> kworker
>>>>>>>> is scheduled by OS randomly, it's difficult to control OS
>>>>>>>> resource
>>>>>>>> and also tune the performance. If target support to use
>>>>>>>> delicated
>>>>>>>> polling task to handle IO, it's useful to control OS resource
>>>>>>>> and
>>>>>>>> gain good performance. So it makes sense to add polling task in
>>>>>>>> rdma-rdma and rdma-tcp modules.
>>>>>>> This is NOT the way to go here.
>>>>>>>
>>>>>>> Both rdma and tcp are driven from workqueue context, which are
>>>>>>> bound
>>>>>>> workqueues.
>>>>>>>
>>>>>>> So there are two ways to go here:
>>>>>>> 1. Add generic port cpuset and use that to direct traffic to the
>>>>>>> appropriate set of cores
>>>>>>> (i.e. select an appropriate comp_vector for rdma and add an
>>>>>>> appropriate
>>>>>>> steering rule
>>>>>>> for tcp).
>>>>>>> 2. Add options to rdma/tcp to use UNBOUND workqueues, and allow
>>>>>>> users
>>>>>>> to
>>>>>>> control
>>>>>>> these UNBOUND workqueues cpumask via sysfs.
>>>>>>>
>>>>>>> (2) will not control interrupts to steer to other workloads
>>>>>>> cpus,
>>>>>>> but
>>>>>>> the handlers may
>>>>>>> run on a set of dedicated cpus.
>>>>>>>
>>>>>>> (1) is a better solution, but harder to implement.
>>>>>>>
>>>>>>> You also should look into nvmet-fc as well (and nvmet-loop for
>>>>>>> that
>>>>>>> matter).
>>>>>> hi Sagi Grimberg,
>>>>>> Thanks for your reply, actually we had tried the first advice you
>>>>>> suggested, but we found the performance was poor when using spdk
>>>>>> as initiator.
>>>>> I suggest that you focus on that instead of what you proposed.
>>>>> What is the source of your poor performance?
>>>> Before these patches, we had used linux's RPS to forward the
>>>> packets
>>>> to a fixed cpu set for nvmet-tcp. But when did that we can still
>>>> not
>>>> cancel the competition between softirq and workqueue since nvme
>>>> target's
>>>> kworker cpu core bind on socket's cpu which is from skb. Besides
>>>> that
>>>> we found workqueue's wait latency was very high even we enabled
>>>> polling
>>>> on nvmet-tcp by module parameter idle_poll_period_usecs. So when
>>>> initiator
>>>> is polling mode, the target of workqueue is the bottleneck. Below
>>>> is
>>>> work's wait latency trace log of our test on our cluster(per node
>>>> uses
>>>> 4 numas 96 cores, 192G memory, one dual ports mellanox CX4LX(25Gbps
>>>> X
>>>> 2)
>>>> ethernet adapter and randrw 1M IO size) by RPS to 6 cpu cores. And
>>>> system's CPU and memory were used about 80%.
>>> I'd try a simple unbound CPU case, steer packets to say cores [0-5]
>>> and
>>> assign
>>> the cpumask of the unbound workqueue to cores [6-11].
>> Okay, thanks for your guide.
>>
>>>> ogden-brown:~ #/usr/share/bcc/tools/wqlat -T -w nvmet_tcp_wq 1 2
>>>> 01:06:59
>>>> usecs : count distribution
>>>> 0 -> 1 : 0 | |
>>>> 2 -> 3 : 0 | |
>>>> 4 -> 7 : 0 | |
>>>> 8 -> 15 : 3 | |
>>>> 16 -> 31 : 10 | |
>>>> 32 -> 63 : 3 | |
>>>> 64 -> 127 : 2 | |
>>>> 128 -> 255 : 0 | |
>>>> 256 -> 511 : 5 | |
>>>> 512 -> 1023 : 12 | |
>>>> 1024 -> 2047 : 26 |* |
>>>> 2048 -> 4095 : 34 |* |
>>>> 4096 -> 8191 : 350 |************ |
>>>> 8192 -> 16383 : 625 |******************************|
>>>> 16384 -> 32767 : 244 |********* |
>>>> 32768 -> 65535 : 39 |* |
>>>>
>>>> 01:07:00
>>>> usecs : count distribution
>>>> 0 -> 1 : 1 | |
>>>> 2 -> 3 : 0 | |
>>>> 4 -> 7 : 4 | |
>>>> 8 -> 15 : 3 | |
>>>> 16 -> 31 : 8 | |
>>>> 32 -> 63 : 10 | |
>>>> 64 -> 127 : 3 | |
>>>> 128 -> 255 : 6 | |
>>>> 256 -> 511 : 8 | |
>>>> 512 -> 1023 : 20 |* |
>>>> 1024 -> 2047 : 19 |* |
>>>> 2048 -> 4095 : 57 |** |
>>>> 4096 -> 8191 : 325 |**************** |
>>>> 8192 -> 16383 : 647 |******************************|
>>>> 16384 -> 32767 : 228 |*********** |
>>>> 32768 -> 65535 : 43 |** |
>>>> 65536 -> 131071 : 1 | |
>>>>
>>>> And the bandwidth of a node is only 3100MB. While we used the patch
>>>> and enable 6 polling task, the bandwidth can be 4000MB. It's a good
>>>> improvement.
>>> I think you will see similar performance with unbound workqueue and
>>> rps.
>> Yes, I remodified the nvmet-tcp/nvmet-rdma code for supporting
>> unbound
>> workqueue, and in same prerequisites of above to run test, and
>> compared
>> the result of unbound workqueue and polling mode task. And I got a
>> good
>> performance for unbound workqueue. For unbound workqueue TCP we got
>> 3850M/node, it's almost equal to polling task. And also tested
>> nvmet-rdma
>> we get 5100M/node for unbound workqueue RDMA versus 5600M for polling
>> task,
>> seems the diff is very small. Anyway, your advice is good.
>
> I'm a bit surprised that you see ~10% delta here. I would look into
> what
> is the root-cause of
> this difference. If indeed the load is high, the overhead of the
> workqueue mgmt should be
> negligible. I'm assuming you used IB_POLL_UNBOUND_WORKQUEUE ?
Yes, we used IB_POLL_UNBOUND_WORKQUEUE to create ib CQ. And I observed
3% CPU
usage of unbound workqueue versus 6% of polling task.
>> Do you think
>> we
>> should submit the unbound workqueue patches for nvmet-tcp and
>> nvmet-rdma
>> to upstream nvmet?
>
> For nvmet-tcp, I think there is merit to split socket processing from
> napi context. For nvmet-rdma
> I think the only difference is if you have multiple CQs assigned with
> the same comp_vector.
>
> How many queues do you have in your test?
We used 24 IO queues to nvmet-rdma target. I think this may also be
related to workqueue's wait latency. We still see some several ms wait
latency for unbound workqueue of RMDA. You can see below trace log.
ogden-brown:~ # /usr/share/bcc/tools/wqlat -T -w ib-comp-unb-wq 1 3
Tracing work queue request latency time... Hit Ctrl-C to end.
10:09:10
usecs : count distribution
0 -> 1 : 6 | |
2 -> 3 : 105 |** |
4 -> 7 : 1732 |******************************|
8 -> 15 : 1597 |******************************|
16 -> 31 : 526 |************ |
32 -> 63 : 543 |************ |
64 -> 127 : 950 |********************* |
128 -> 255 : 1335 |***************************** |
256 -> 511 : 1534 |******************************|
512 -> 1023 : 1039 |*********************** |
1024 -> 2047 : 592 |************* |
2048 -> 4095 : 112 |** |
4096 -> 8191 : 6 | |
10:09:11
usecs : count distribution
0 -> 1 : 3 | |
2 -> 3 : 62 |* |
4 -> 7 : 1459 |***************************** |
8 -> 15 : 1869 |******************************|
16 -> 31 : 612 |************* |
32 -> 63 : 478 |********** |
64 -> 127 : 844 |****************** |
128 -> 255 : 1123 |************************ |
256 -> 511 : 1278 |*************************** |
512 -> 1023 : 1113 |*********************** |
1024 -> 2047 : 632 |************* |
2048 -> 4095 : 158 |*** |
4096 -> 8191 : 18 | |
8192 -> 16383 : 1 | |
10:09:12
usecs : count distribution
0 -> 1 : 1 | |
2 -> 3 : 68 |* |
4 -> 7 : 1399 |*************************** |
8 -> 15 : 1822 |******************************|
16 -> 31 : 559 |************ |
32 -> 63 : 513 |*********** |
64 -> 127 : 906 |******************* |
128 -> 255 : 1217 |*********************** |
256 -> 511 : 1391 |*************************** |
512 -> 1023 : 1135 |************************ |
1024 -> 2047 : 569 |************ |
2048 -> 4095 : 110 |** |
4096 -> 8191 : 26 | |
8192 -> 16383 : 11 | |
Thanks,
Ping
On 7/4/24 13:35, Ping Gan wrote: >> On 7/4/24 11:10, Ping Gan wrote: >>>> On 02/07/2024 13:02, Ping Gan wrote: >>>>>> On 01/07/2024 10:42, Ping Gan wrote: >>>>>>>> Hey Ping Gan, >>>>>>>> >>>>>>>> >>>>>>>> On 26/06/2024 11:28, Ping Gan wrote: >>>>>>>>> When running nvmf on SMP platform, current nvme target's RDMA >>>>>>>>> and >>>>>>>>> TCP use kworker to handle IO. But if there is other high >>>>>>>>> workload >>>>>>>>> in the system(eg: on kubernetes), the competition between the >>>>>>>>> kworker and other workload is very radical. And since the >>>>>>>>> kworker >>>>>>>>> is scheduled by OS randomly, it's difficult to control OS >>>>>>>>> resource >>>>>>>>> and also tune the performance. If target support to use >>>>>>>>> delicated >>>>>>>>> polling task to handle IO, it's useful to control OS resource >>>>>>>>> and >>>>>>>>> gain good performance. So it makes sense to add polling task in >>>>>>>>> rdma-rdma and rdma-tcp modules. >>>>>>>> This is NOT the way to go here. >>>>>>>> >>>>>>>> Both rdma and tcp are driven from workqueue context, which are >>>>>>>> bound >>>>>>>> workqueues. >>>>>>>> >>>>>>>> So there are two ways to go here: >>>>>>>> 1. Add generic port cpuset and use that to direct traffic to the >>>>>>>> appropriate set of cores >>>>>>>> (i.e. select an appropriate comp_vector for rdma and add an >>>>>>>> appropriate >>>>>>>> steering rule >>>>>>>> for tcp). >>>>>>>> 2. Add options to rdma/tcp to use UNBOUND workqueues, and allow >>>>>>>> users >>>>>>>> to >>>>>>>> control >>>>>>>> these UNBOUND workqueues cpumask via sysfs. >>>>>>>> >>>>>>>> (2) will not control interrupts to steer to other workloads >>>>>>>> cpus, >>>>>>>> but >>>>>>>> the handlers may >>>>>>>> run on a set of dedicated cpus. >>>>>>>> >>>>>>>> (1) is a better solution, but harder to implement. >>>>>>>> >>>>>>>> You also should look into nvmet-fc as well (and nvmet-loop for >>>>>>>> that >>>>>>>> matter). >>>>>>> hi Sagi Grimberg, >>>>>>> Thanks for your reply, actually we had tried the first advice you >>>>>>> suggested, but we found the performance was poor when using spdk >>>>>>> as initiator. >>>>>> I suggest that you focus on that instead of what you proposed. >>>>>> What is the source of your poor performance? >>>>> Before these patches, we had used linux's RPS to forward the >>>>> packets >>>>> to a fixed cpu set for nvmet-tcp. But when did that we can still >>>>> not >>>>> cancel the competition between softirq and workqueue since nvme >>>>> target's >>>>> kworker cpu core bind on socket's cpu which is from skb. Besides >>>>> that >>>>> we found workqueue's wait latency was very high even we enabled >>>>> polling >>>>> on nvmet-tcp by module parameter idle_poll_period_usecs. So when >>>>> initiator >>>>> is polling mode, the target of workqueue is the bottleneck. Below >>>>> is >>>>> work's wait latency trace log of our test on our cluster(per node >>>>> uses >>>>> 4 numas 96 cores, 192G memory, one dual ports mellanox CX4LX(25Gbps >>>>> X >>>>> 2) >>>>> ethernet adapter and randrw 1M IO size) by RPS to 6 cpu cores. And >>>>> system's CPU and memory were used about 80%. >>>> I'd try a simple unbound CPU case, steer packets to say cores [0-5] >>>> and >>>> assign >>>> the cpumask of the unbound workqueue to cores [6-11]. >>> Okay, thanks for your guide. >>> >>>>> ogden-brown:~ #/usr/share/bcc/tools/wqlat -T -w nvmet_tcp_wq 1 2 >>>>> 01:06:59 >>>>> usecs : count distribution >>>>> 0 -> 1 : 0 | | >>>>> 2 -> 3 : 0 | | >>>>> 4 -> 7 : 0 | | >>>>> 8 -> 15 : 3 | | >>>>> 16 -> 31 : 10 | | >>>>> 32 -> 63 : 3 | | >>>>> 64 -> 127 : 2 | | >>>>> 128 -> 255 : 0 | | >>>>> 256 -> 511 : 5 | | >>>>> 512 -> 1023 : 12 | | >>>>> 1024 -> 2047 : 26 |* | >>>>> 2048 -> 4095 : 34 |* | >>>>> 4096 -> 8191 : 350 |************ | >>>>> 8192 -> 16383 : 625 |******************************| >>>>> 16384 -> 32767 : 244 |********* | >>>>> 32768 -> 65535 : 39 |* | >>>>> >>>>> 01:07:00 >>>>> usecs : count distribution >>>>> 0 -> 1 : 1 | | >>>>> 2 -> 3 : 0 | | >>>>> 4 -> 7 : 4 | | >>>>> 8 -> 15 : 3 | | >>>>> 16 -> 31 : 8 | | >>>>> 32 -> 63 : 10 | | >>>>> 64 -> 127 : 3 | | >>>>> 128 -> 255 : 6 | | >>>>> 256 -> 511 : 8 | | >>>>> 512 -> 1023 : 20 |* | >>>>> 1024 -> 2047 : 19 |* | >>>>> 2048 -> 4095 : 57 |** | >>>>> 4096 -> 8191 : 325 |**************** | >>>>> 8192 -> 16383 : 647 |******************************| >>>>> 16384 -> 32767 : 228 |*********** | >>>>> 32768 -> 65535 : 43 |** | >>>>> 65536 -> 131071 : 1 | | >>>>> >>>>> And the bandwidth of a node is only 3100MB. While we used the patch >>>>> and enable 6 polling task, the bandwidth can be 4000MB. It's a good >>>>> improvement. >>>> I think you will see similar performance with unbound workqueue and >>>> rps. >>> Yes, I remodified the nvmet-tcp/nvmet-rdma code for supporting >>> unbound >>> workqueue, and in same prerequisites of above to run test, and >>> compared >>> the result of unbound workqueue and polling mode task. And I got a >>> good >>> performance for unbound workqueue. For unbound workqueue TCP we got >>> 3850M/node, it's almost equal to polling task. And also tested >>> nvmet-rdma >>> we get 5100M/node for unbound workqueue RDMA versus 5600M for polling >>> task, >>> seems the diff is very small. Anyway, your advice is good. >> I'm a bit surprised that you see ~10% delta here. I would look into >> what >> is the root-cause of >> this difference. If indeed the load is high, the overhead of the >> workqueue mgmt should be >> negligible. I'm assuming you used IB_POLL_UNBOUND_WORKQUEUE ? > Yes, we used IB_POLL_UNBOUND_WORKQUEUE to create ib CQ. And I observed > 3% CPU > usage of unbound workqueue versus 6% of polling task. > >>> Do you think >>> we >>> should submit the unbound workqueue patches for nvmet-tcp and >>> nvmet-rdma >>> to upstream nvmet? >> For nvmet-tcp, I think there is merit to split socket processing from >> napi context. For nvmet-rdma >> I think the only difference is if you have multiple CQs assigned with >> the same comp_vector. >> >> How many queues do you have in your test? > We used 24 IO queues to nvmet-rdma target. I think this may also be > related to workqueue's wait latency. We still see some several ms wait > latency for unbound workqueue of RMDA. You can see below trace log. What is the queue size of each? what rdma device are you using?
> On 7/4/24 13:35, Ping Gan wrote: >>> On 7/4/24 11:10, Ping Gan wrote: >>>>> On 02/07/2024 13:02, Ping Gan wrote: >>>>>>> On 01/07/2024 10:42, Ping Gan wrote: >>>>>>>>> Hey Ping Gan, >>>>>>>>> >>>>>>>>> >>>>>>>>> On 26/06/2024 11:28, Ping Gan wrote: >>>>>>>>>> When running nvmf on SMP platform, current nvme target's RDMA >>>>>>>>>> and >>>>>>>>>> TCP use kworker to handle IO. But if there is other high >>>>>>>>>> workload >>>>>>>>>> in the system(eg: on kubernetes), the competition between the >>>>>>>>>> kworker and other workload is very radical. And since the >>>>>>>>>> kworker >>>>>>>>>> is scheduled by OS randomly, it's difficult to control OS >>>>>>>>>> resource >>>>>>>>>> and also tune the performance. If target support to use >>>>>>>>>> delicated >>>>>>>>>> polling task to handle IO, it's useful to control OS resource >>>>>>>>>> and >>>>>>>>>> gain good performance. So it makes sense to add polling task >>>>>>>>>> in >>>>>>>>>> rdma-rdma and rdma-tcp modules. >>>>>>>>> This is NOT the way to go here. >>>>>>>>> >>>>>>>>> Both rdma and tcp are driven from workqueue context, which are >>>>>>>>> bound >>>>>>>>> workqueues. >>>>>>>>> >>>>>>>>> So there are two ways to go here: >>>>>>>>> 1. Add generic port cpuset and use that to direct traffic to >>>>>>>>> the >>>>>>>>> appropriate set of cores >>>>>>>>> (i.e. select an appropriate comp_vector for rdma and add an >>>>>>>>> appropriate >>>>>>>>> steering rule >>>>>>>>> for tcp). >>>>>>>>> 2. Add options to rdma/tcp to use UNBOUND workqueues, and >>>>>>>>> allow >>>>>>>>> users >>>>>>>>> to >>>>>>>>> control >>>>>>>>> these UNBOUND workqueues cpumask via sysfs. >>>>>>>>> >>>>>>>>> (2) will not control interrupts to steer to other workloads >>>>>>>>> cpus, >>>>>>>>> but >>>>>>>>> the handlers may >>>>>>>>> run on a set of dedicated cpus. >>>>>>>>> >>>>>>>>> (1) is a better solution, but harder to implement. >>>>>>>>> >>>>>>>>> You also should look into nvmet-fc as well (and nvmet-loop for >>>>>>>>> that >>>>>>>>> matter). >>>>>>>> hi Sagi Grimberg, >>>>>>>> Thanks for your reply, actually we had tried the first advice >>>>>>>> you >>>>>>>> suggested, but we found the performance was poor when using >>>>>>>> spdk >>>>>>>> as initiator. >>>>>>> I suggest that you focus on that instead of what you proposed. >>>>>>> What is the source of your poor performance? >>>>>> Before these patches, we had used linux's RPS to forward the >>>>>> packets >>>>>> to a fixed cpu set for nvmet-tcp. But when did that we can still >>>>>> not >>>>>> cancel the competition between softirq and workqueue since nvme >>>>>> target's >>>>>> kworker cpu core bind on socket's cpu which is from skb. Besides >>>>>> that >>>>>> we found workqueue's wait latency was very high even we enabled >>>>>> polling >>>>>> on nvmet-tcp by module parameter idle_poll_period_usecs. So when >>>>>> initiator >>>>>> is polling mode, the target of workqueue is the bottleneck. Below >>>>>> is >>>>>> work's wait latency trace log of our test on our cluster(per node >>>>>> uses >>>>>> 4 numas 96 cores, 192G memory, one dual ports mellanox >>>>>> CX4LX(25Gbps >>>>>> X >>>>>> 2) >>>>>> ethernet adapter and randrw 1M IO size) by RPS to 6 cpu cores. >>>>>> And >>>>>> system's CPU and memory were used about 80%. >>>>> I'd try a simple unbound CPU case, steer packets to say cores >>>>> [0-5] >>>>> and >>>>> assign >>>>> the cpumask of the unbound workqueue to cores [6-11]. >>>> Okay, thanks for your guide. >>>> >>>>>> ogden-brown:~ #/usr/share/bcc/tools/wqlat -T -w nvmet_tcp_wq 1 2 >>>>>> 01:06:59 >>>>>> usecs : count distribution >>>>>> 0 -> 1 : 0 | | >>>>>> 2 -> 3 : 0 | | >>>>>> 4 -> 7 : 0 | | >>>>>> 8 -> 15 : 3 | | >>>>>> 16 -> 31 : 10 | | >>>>>> 32 -> 63 : 3 | | >>>>>> 64 -> 127 : 2 | | >>>>>> 128 -> 255 : 0 | | >>>>>> 256 -> 511 : 5 | | >>>>>> 512 -> 1023 : 12 | | >>>>>> 1024 -> 2047 : 26 |* | >>>>>> 2048 -> 4095 : 34 |* | >>>>>> 4096 -> 8191 : 350 |************ | >>>>>> 8192 -> 16383 : 625 |******************************| >>>>>> 16384 -> 32767 : 244 |********* | >>>>>> 32768 -> 65535 : 39 |* | >>>>>> >>>>>> 01:07:00 >>>>>> usecs : count distribution >>>>>> 0 -> 1 : 1 | | >>>>>> 2 -> 3 : 0 | | >>>>>> 4 -> 7 : 4 | | >>>>>> 8 -> 15 : 3 | | >>>>>> 16 -> 31 : 8 | | >>>>>> 32 -> 63 : 10 | | >>>>>> 64 -> 127 : 3 | | >>>>>> 128 -> 255 : 6 | | >>>>>> 256 -> 511 : 8 | | >>>>>> 512 -> 1023 : 20 |* | >>>>>> 1024 -> 2047 : 19 |* | >>>>>> 2048 -> 4095 : 57 |** | >>>>>> 4096 -> 8191 : 325 |**************** | >>>>>> 8192 -> 16383 : 647 |******************************| >>>>>> 16384 -> 32767 : 228 |*********** | >>>>>> 32768 -> 65535 : 43 |** | >>>>>> 65536 -> 131071 : 1 | | >>>>>> >>>>>> And the bandwidth of a node is only 3100MB. While we used the >>>>>> patch >>>>>> and enable 6 polling task, the bandwidth can be 4000MB. It's a >>>>>> good >>>>>> improvement. >>>>> I think you will see similar performance with unbound workqueue >>>>> and >>>>> rps. >>>> Yes, I remodified the nvmet-tcp/nvmet-rdma code for supporting >>>> unbound >>>> workqueue, and in same prerequisites of above to run test, and >>>> compared >>>> the result of unbound workqueue and polling mode task. And I got a >>>> good >>>> performance for unbound workqueue. For unbound workqueue TCP we got >>>> 3850M/node, it's almost equal to polling task. And also tested >>>> nvmet-rdma >>>> we get 5100M/node for unbound workqueue RDMA versus 5600M for >>>> polling >>>> task, >>>> seems the diff is very small. Anyway, your advice is good. >>> I'm a bit surprised that you see ~10% delta here. I would look into >>> what >>> is the root-cause of >>> this difference. If indeed the load is high, the overhead of the >>> workqueue mgmt should be >>> negligible. I'm assuming you used IB_POLL_UNBOUND_WORKQUEUE ? >> Yes, we used IB_POLL_UNBOUND_WORKQUEUE to create ib CQ. And I >> observed >> 3% CPU >> usage of unbound workqueue versus 6% of polling task. >> >>>> Do you think >>>> we >>>> should submit the unbound workqueue patches for nvmet-tcp and >>>> nvmet-rdma >>>> to upstream nvmet? >>> For nvmet-tcp, I think there is merit to split socket processing >>> from >>> napi context. For nvmet-rdma >>> I think the only difference is if you have multiple CQs assigned >>> with >>> the same comp_vector. >>> >>> How many queues do you have in your test? >> We used 24 IO queues to nvmet-rdma target. I think this may also be >> related to workqueue's wait latency. We still see some several ms >> wait >> latency for unbound workqueue of RMDA. You can see below trace log. > > What is the queue size of each? what rdma device are you using? All the queue's IO size is 1M and queue depth is 32. The rdma deive is Mellanox CX4LX dual ports bonding. And in poll task we used IB_POLL_DIRECT to create CQ versus IB_POLL_UNBOUND_WORKQUEUE for workqueue. Thanks, Ping
When running nvmf on SMP platform, current nvme target's RDMA and TCP use kworker to handle IO. But if there is other high workload in the system(eg: on kubernetes), the competition between the kworker and other workload is very radical. And since the kworker is scheduled by OS randomly, it's difficult to control OS resource and also tune the performance. If target support to use delicated polling task to handle IO, it's useful to control OS resource and gain good performance. So it makes sense to add polling task in rdma-rdma and rdma-tcp modules. Ping Gan (2): nvmet-rdma: add polling cq task for nvmet-rdma nvmet-tcp: add polling task for nvmet-tcp drivers/nvme/target/rdma.c | 331 +++++++++++++++++++++++++++++++++- drivers/nvme/target/tcp.c | 356 +++++++++++++++++++++++++++++++++++-- 2 files changed, 665 insertions(+), 22 deletions(-) -- 2.26.2
When running nvmf on SMP platform, current nvme target's RDMA and TCP use kworker to handle IO. But if there is other high workload in the system(eg: on kubernetes), the competition between the kworker and other workload is very radical. And since the kworker is scheduled by OS randomly, it's difficult to control OS resource and also tune the performance. If target support to use delicated polling task to handle IO, it's useful to control OS resource and gain good performance. So it makes sense to add polling task in rdma-rdma and rdma-tcp modules. Ping Gan (2): nvmet-rdma: add polling cq task for nvmet-rdma nvmet-tcp: add polling task for nvmet-tcp drivers/nvme/target/rdma.c | 331 +++++++++++++++++++++++++++++++++- drivers/nvme/target/tcp.c | 356 +++++++++++++++++++++++++++++++++++-- 2 files changed, 665 insertions(+), 22 deletions(-) -- 2.26.2
© 2016 - 2026 Red Hat, Inc.