Previously posted series patches:
v1 - http://lists.nongnu.org/archive/html/qemu-devel/2017-03/msg02044.html
v2 - http://lists.nongnu.org/archive/html/qemu-devel/2017-03/msg05080.html
v3 - http://lists.nongnu.org/archive/html/qemu-devel/2017-04/msg00074.html
This series helps to optimize the I/O performance of VMDK driver.
Patch 1 helps us to move vmdk_find_offset_in_cluster.
Patch 2 & 3 perform a simple function re-naming tasks.
Patch 4 is used to factor out metadata loading code and implement it in separate
functions. This will help us to avoid code duplication in future patches of this
series.
Patch 5 helps to set the upper limit of the bytes handled in one cycle.
Patch 6 adds new functions to help us allocate multiple clusters according to
the size requested, perform COW if required and return the offset of the first
newly allocated cluster.
Patch 7 changes the metadata update code to update the L2 tables for multiple
clusters at once.
Patch 8 helps us to finally change vmdk_get_cluster_offset() to find cluster offset
only as cluster allocation task is now handled by vmdk_alloc_clusters()
Note: v4 has an addition of new optimization of calling bdrv_pwrite_sync() only
once for atmost 512 clusters, as a result performance has increased to a great
extent (earlier till v2 it was 29%).
Optimization test results:
This patch series improves 128 KB sequential write performance to an
empty VMDK file by 54%
Benchmark command: ./qemu-img bench -w -c 1024 -s 128K -d 1 -t none -f
vmdk test.vmdk
As showoff, these patches now complete a 128M write request on an empty VMDK file
on my slow laptop in just 2.7 secs compared to 3.5 mins as in v2.9.0
This is obviously using qemu-io with "--cache writeback" and not an official
benchmark, but worth mentioning from a newbie's perspective.
Note: These patches pass all 41/41 tests suitable for the VMDK driver.
Changes in v4:
- fix commit message in patch 1 (fam)
- drop size_to_clusters() function (fam)
- fix grammatical errors in function documentations (fam)
- factor out metadata loading coding in a separate patch (patch 4) (fam)
- rename vmdk_alloc_cluster_offset() to vmdk_alloc_clusters() (fam)
- break patch 4(in v3) into separate patches (patch 3 and 8) (fam)
- rename extent_size to extent_end (fam)
- use QEMU_ALIGN_UP instead of vmdk_align_offset. (fam)
- drop next and simply do m_data = m_data->next (fam)
Changes in v3:
- move size_to_clusters() from patch 1 to 3 (fam)
- use DIV_ROUND_UP in size_to_clusters (fam)
- make patch 2 compilable (fam)
- rename vmdk_L2update as vmdk_l2update and use UINT32_MAX (fam)
- combine patch 3 and patch 4 (as in v2) to make them compilable (fam)
- call bdrv_pwrite_sync() for batches of atmost 512 clusters at once (fam)
Changes in v2:
- segregate the ugly Patch 1 in v1 into 6 readable and sensible patches
- include benchmark test results in v2
Ashijeet Acharya (8):
vmdk: Move vmdk_find_offset_in_cluster() to the top
vmdk: Rename get_whole_cluster() to vmdk_perform_cow()
vmdk: Rename get_cluster_offset() to vmdk_get_cluster_offset()
vmdk: Factor out metadata loading code out of
vmdk_get_cluster_offset()
vmdk: Set maximum bytes allocated in one cycle
vmdk: New functions to assist allocating multiple clusters
vmdk: Update metadata for multiple clusters
vmdk: Make vmdk_get_cluster_offset() return cluster offset only
block/vmdk.c | 530 +++++++++++++++++++++++++++++++++++++++++++++--------------
1 file changed, 408 insertions(+), 122 deletions(-)
--
2.6.2