hw/9pfs/9p.c | 128 +++++++++++++++++++++++------------ hw/9pfs/9p.h | 2 +- tests/qtest/virtio-9p-test.c | 30 ++++++++ 3 files changed, 116 insertions(+), 44 deletions(-)
This patch set reduces latency of Twalk handling by reducing the amount of
thread hops, similar to previous Treaddir optimization.
The performance gain of this patch set is not as spectacular as previously
with Treaddir, but there is definitely a measurable difference.
The actualy performance optimization is patch 7. With the benchmark of
patch 2, the runtime of the Twalk test was cut in half. In real world tests
I measured a performance gain (i.e. running an entire guest OS, and hence
mixed with all other kinds of 9p requests) of about 2%, again measured in a
mix, not concentrated on Twalk alone.
v1 -> v2:
* Added a separate [NEW PATCH 1] for fixing not_same_qid() semantic.
* Code style: use sizeof(var-name) instead of sizeof(type-name)
[PATCH 3].
* Split out dropping fid_to_qid() into separate [PATCH 4].
* Split out replacing not_same_qid() by same_stat_id() into separate
[PATCH 5].
* Split out dropping root_qid into separate [PATCH 6].
Christian Schoenebeck (7):
9pfs: fix not_same_qid()
9pfs: Twalk benchmark
9pfs: capture root stat
9pfs: drop fid_to_qid()
9pfs: replace not_same_qid() by same_stat_id()
9pfs: drop root_qid
9pfs: reduce latency of Twalk
hw/9pfs/9p.c | 128 +++++++++++++++++++++++------------
hw/9pfs/9p.h | 2 +-
tests/qtest/virtio-9p-test.c | 30 ++++++++
3 files changed, 116 insertions(+), 44 deletions(-)
--
2.20.1
On Fri, 4 Jun 2021 17:48:49 +0200 Christian Schoenebeck <qemu_oss@crudebyte.com> wrote: > This patch set reduces latency of Twalk handling by reducing the amount of > thread hops, similar to previous Treaddir optimization. > > The performance gain of this patch set is not as spectacular as previously > with Treaddir, but there is definitely a measurable difference. > > The actualy performance optimization is patch 7. With the benchmark of > patch 2, the runtime of the Twalk test was cut in half. In real world tests > I measured a performance gain (i.e. running an entire guest OS, and hence > mixed with all other kinds of 9p requests) of about 2%, again measured in a > mix, not concentrated on Twalk alone. > > v1 -> v2: > > * Added a separate [NEW PATCH 1] for fixing not_same_qid() semantic. > Strangely, patch 1 appears between patch 3 and 4 in my email client. > * Code style: use sizeof(var-name) instead of sizeof(type-name) > [PATCH 3]. > > * Split out dropping fid_to_qid() into separate [PATCH 4]. > > * Split out replacing not_same_qid() by same_stat_id() into separate > [PATCH 5]. > > * Split out dropping root_qid into separate [PATCH 6]. > I could have a look at all the patches except the last one. LGTM. You can add my R-b for patches 1 and 3 to 6. Reviewed-by: Greg Kurz <groug@kaod.org> I'll try to find some time for patch 7 next week. > Christian Schoenebeck (7): > 9pfs: fix not_same_qid() > 9pfs: Twalk benchmark > 9pfs: capture root stat > 9pfs: drop fid_to_qid() > 9pfs: replace not_same_qid() by same_stat_id() > 9pfs: drop root_qid > 9pfs: reduce latency of Twalk > > hw/9pfs/9p.c | 128 +++++++++++++++++++++++------------ > hw/9pfs/9p.h | 2 +- > tests/qtest/virtio-9p-test.c | 30 ++++++++ > 3 files changed, 116 insertions(+), 44 deletions(-) >
On Freitag, 4. Juni 2021 18:31:28 CEST Greg Kurz wrote: > On Fri, 4 Jun 2021 17:48:49 +0200 > > Christian Schoenebeck <qemu_oss@crudebyte.com> wrote: > > This patch set reduces latency of Twalk handling by reducing the amount of > > thread hops, similar to previous Treaddir optimization. > > > > The performance gain of this patch set is not as spectacular as previously > > with Treaddir, but there is definitely a measurable difference. > > > > The actualy performance optimization is patch 7. With the benchmark of > > patch 2, the runtime of the Twalk test was cut in half. In real world > > tests > > I measured a performance gain (i.e. running an entire guest OS, and hence > > mixed with all other kinds of 9p requests) of about 2%, again measured in > > a > > mix, not concentrated on Twalk alone. > > > > v1 -> v2: > > * Added a separate [NEW PATCH 1] for fixing not_same_qid() semantic. > > Strangely, patch 1 appears between patch 3 and 4 in my email client. My bad, I forgot to take care about the time stamps of the individual patches this time. > > * Code style: use sizeof(var-name) instead of sizeof(type-name) > > > > [PATCH 3]. > > > > * Split out dropping fid_to_qid() into separate [PATCH 4]. > > > > * Split out replacing not_same_qid() by same_stat_id() into separate > > > > [PATCH 5]. > > > > * Split out dropping root_qid into separate [PATCH 6]. > > I could have a look at all the patches except the last one. LGTM. > You can add my R-b for patches 1 and 3 to 6. > > Reviewed-by: Greg Kurz <groug@kaod.org> Queued patches 1, 3..6 on 9p.next: https://github.com/cschoenebeck/qemu/commits/9p.next Thanks! > I'll try to find some time for patch 7 next week. > Twalk can be tricky, so no hurry! It takes what it takes. > > Christian Schoenebeck (7): > > 9pfs: fix not_same_qid() > > 9pfs: Twalk benchmark > > 9pfs: capture root stat > > 9pfs: drop fid_to_qid() > > 9pfs: replace not_same_qid() by same_stat_id() > > 9pfs: drop root_qid > > 9pfs: reduce latency of Twalk > > > > hw/9pfs/9p.c | 128 +++++++++++++++++++++++------------ > > hw/9pfs/9p.h | 2 +- > > tests/qtest/virtio-9p-test.c | 30 ++++++++ > > 3 files changed, 116 insertions(+), 44 deletions(-) Best regards, Christian Schoenebeck
On Freitag, 4. Juni 2021 20:23:21 CEST Christian Schoenebeck wrote: > On Freitag, 4. Juni 2021 18:31:28 CEST Greg Kurz wrote: > > On Fri, 4 Jun 2021 17:48:49 +0200 > > > > Christian Schoenebeck <qemu_oss@crudebyte.com> wrote: > > > This patch set reduces latency of Twalk handling by reducing the amount > > > of > > > thread hops, similar to previous Treaddir optimization. > > > > > > The performance gain of this patch set is not as spectacular as > > > previously > > > with Treaddir, but there is definitely a measurable difference. > > > > > > The actualy performance optimization is patch 7. With the benchmark of > > > patch 2, the runtime of the Twalk test was cut in half. In real world > > > tests > > > I measured a performance gain (i.e. running an entire guest OS, and > > > hence > > > mixed with all other kinds of 9p requests) of about 2%, again measured > > > in > > > a > > > mix, not concentrated on Twalk alone. > > > > > > v1 -> v2: > > > * Added a separate [NEW PATCH 1] for fixing not_same_qid() semantic. > > > > Strangely, patch 1 appears between patch 3 and 4 in my email client. > > My bad, I forgot to take care about the time stamps of the individual > patches this time. > > > > * Code style: use sizeof(var-name) instead of sizeof(type-name) > > > > > > [PATCH 3]. > > > > > > * Split out dropping fid_to_qid() into separate [PATCH 4]. > > > > > > * Split out replacing not_same_qid() by same_stat_id() into separate > > > > > > [PATCH 5]. > > > > > > * Split out dropping root_qid into separate [PATCH 6]. > > > > I could have a look at all the patches except the last one. LGTM. > > You can add my R-b for patches 1 and 3 to 6. > > > > Reviewed-by: Greg Kurz <groug@kaod.org> > > Queued patches 1, 3..6 on 9p.next: > > https://github.com/cschoenebeck/qemu/commits/9p.next > > Thanks! > > > I'll try to find some time for patch 7 next week. > > Twalk can be tricky, so no hurry! It takes what it takes. > > > > Christian Schoenebeck (7): > > > 9pfs: fix not_same_qid() > > > 9pfs: Twalk benchmark > > > 9pfs: capture root stat > > > 9pfs: drop fid_to_qid() > > > 9pfs: replace not_same_qid() by same_stat_id() > > > 9pfs: drop root_qid > > > 9pfs: reduce latency of Twalk > > > > > > hw/9pfs/9p.c | 128 +++++++++++++++++++++++------------ > > > hw/9pfs/9p.h | 2 +- > > > tests/qtest/virtio-9p-test.c | 30 ++++++++ > > > 3 files changed, 116 insertions(+), 44 deletions(-) Ping How is your schedule Greg? Any chance for a time slice on this one? Best regards, Christian Schoenebeck
© 2016 - 2024 Red Hat, Inc.