This patch set adds support for native vectored AIO ops. Instead of using our
own thread pool for queueing io ops, we use the kernel.
Numbers - short story: Overall increase of at least 10%, with sequential
ops benefiting the most.
Numbers - long story:
Before:
Version 1.96 ------Sequential Output------ --Sequential Input- --Random-
Concurrency 2 -Per Chr- --Block-- -Rewrite- -Per Chr- --Block-- --Seeks--
Machine Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP /sec %CP
512 1G 550 97 26638 7 25050 4 2942 91 339188 35 1711 68
Latency 40602us 12269ms 2755ms 27333us 119ms 77816us
Version 1.96 ------Sequential Create------ --------Random Create--------
512 -Create-- --Read--- -Delete-- -Create-- --Read--- -Delete--
files /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP
16 12172 22 +++++ +++ 23771 34 20209 36 +++++ +++ 22181 33
Latency 153ms 566us 571us 204us 56us 624us
1.96,1.96,512,2,1319996851,1G,,550,97,26638,7,25050,4,2942,91,339188,35,1711,68,16,,,,,12172,22,+++++,+++,23771,34,20209,36,+++++,+++,22181,33,40602us,12269ms,2755ms,27333us,119ms,77816us,153ms,566us,571us,204us,56us,624us
After:
Version 1.96 ------Sequential Output------ --Sequential Input- --Random-
Concurrency 2 -Per Chr- --Block-- -Rewrite- -Per Chr- --Block-- --Seeks--
Machine Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP /sec %CP
tux 512M 565 96 49307 19 33464 6 3270 99 923539 57 2599 64
Latency 15285us 240ms 2496ms 7478us 5586us 2878ms
Version 1.96 ------Sequential Create------ --------Random Create--------
tux -Create-- --Read--- -Delete-- -Create-- --Read--- -Delete--
files /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP
16 26992 48 +++++ +++ 27018 38 25618 45 +++++ +++ 24272 34
Latency 1145us 1311us 1708us 408us 186us 283us
1.96,1.96,tux,2,1319997274,512M,,565,96,49307,19,33464,6,3270,99,923539,57,2599,64,16,,,,,26992,48,+++++,+++,27018,38,25618,45,+++++,+++,24272,34,15285us,240ms,2496ms,7478us,5586us,2878ms,1145us,1311us,1708us,408us,186us,283us
I'd be happy if someone could run more benchmarks, as these ones looks very
optimistic, but I kept getting similar results over and over.
Sasha Levin (11):
kvm tools: Switch to using an enum for disk image types
kvm tools: Hold a copy of ops struct inside disk_image
kvm tools: Remove the non-iov interface from disk image ops
kvm tools: Modify disk ops usage
kvm tools: Modify behaviour on missing ops ptr
kvm tools: Add optional callback on disk op completion
kvm tools: Remove qcow nowrite function
kvm tools: Split io request from completion
kvm tools: Hook virtio-blk completion to disk op completion
kvm tools: Add aio read write functions
kvm tools: Use native vectored AIO in virtio-blk
tools/kvm/Makefile | 1 +
tools/kvm/disk/core.c | 116 +++++++++++++++++++---------------
tools/kvm/disk/qcow.c | 58 ++++++++++++++---
tools/kvm/disk/raw.c | 87 +++++++++++++++++--------
tools/kvm/include/kvm/disk-image.h | 50 ++++++++-------
tools/kvm/include/kvm/read-write.h | 6 ++
tools/kvm/include/kvm/virtio-blk.h | 1 +
tools/kvm/read-write.c | 24 +++++++
tools/kvm/virtio/blk.c | 122 ++++++++++++++++++++++++------------
9 files changed, 312 insertions(+), 153 deletions(-)
--
1.7.7.1
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
