| Message ID | 1595475001-90945-1-git-send-email-feng.tang@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [RFC] makefile: add debug option to enable function aligned on 32 bytes | expand |
On Thu, 23 Jul 2020 11:30:01 +0800 Feng Tang <feng.tang@intel.com> wrote: > Recently 0day reported many strange performance changes (regression > or improvement), in which there was no obvious relation between > the culprit commit and the benchmark at the first look, and it causes > people to doubt the test itself is wrong. > > Upon further check, many of these cases are caused by the change > to the alignment of kernel text or data, as whole text/data of kernel > are linked together, change in one domain may affect alignments of > other domains. > > gcc has an option '-falign-functions=n' to force text aligned, and with > that option enabled, some of those performance changes will be gone, > like [1][2][3]. > > Add this option so that developers and 0day can easily find performance > bump caused by text alignment change, Would they use it this way, or would they simply always enable the option to reduce the variability? It makes sense, but is it actually known that this does reduce the variability? > as tracking these strange bump > is quite time consuming. Though it can't help in other cases like data > alignment changes like [4]. > > Following is some size data for v5.7 kernel built with a RHEL config > used in 0day: > > text data bss dec filename > 19738771 13292906 5554236 38585913 vmlinux.noalign > 19758591 13297002 5529660 38585253 vmlinux.align32 > > Raw vmlinux size in bytes: > > v5.7 v5.7+align32 > 253950832 254018000 +0.02% > > Some benchmark data, most of them have no big change: > > * hackbench: [ -1.8%, +0.5%] > > * fsmark: [ -3.2%, +3.4%] # ext4/xfs/btrfs > > * kbuild: [ -2.0%, +0.9%] > > * will-it-scale: [ -0.5%, +1.8%] # mmap1/pagefault3 > > * netperf: > - TCP_CRR [+16.6%, +97.4%] > - TCP_RR [-18.5%, -1.8%] > - TCP_STREAM [ -1.1%, +1.9%] What do the numbers in [] mean? The TCP_CRR results look remarkable? > [1] https://lore.kernel.org/lkml/20200114085637.GA29297@shao2-debian/ > [2] https://lore.kernel.org/lkml/20200330011254.GA14393@feng-iot/ > [3] https://lore.kernel.org/lkml/1d98d1f0-fe84-6df7-f5bd-f4cb2cdb7f45@intel.com/ > [4] https://lore.kernel.org/lkml/20200205123216.GO12867@shao2-debian/ >
Hi Andrew, Thanks for the review. On Wed, Jul 22, 2020 at 08:39:19PM -0700, Andrew Morton wrote: > On Thu, 23 Jul 2020 11:30:01 +0800 Feng Tang <feng.tang@intel.com> wrote: > > > Recently 0day reported many strange performance changes (regression > > or improvement), in which there was no obvious relation between > > the culprit commit and the benchmark at the first look, and it causes > > people to doubt the test itself is wrong. > > > > Upon further check, many of these cases are caused by the change > > to the alignment of kernel text or data, as whole text/data of kernel > > are linked together, change in one domain may affect alignments of > > other domains. > > > > gcc has an option '-falign-functions=n' to force text aligned, and with > > that option enabled, some of those performance changes will be gone, > > like [1][2][3]. > > > > Add this option so that developers and 0day can easily find performance > > bump caused by text alignment change, > > Would they use it this way, or would they simply always enable the > option to reduce the variability We've had concerns about side effects, like increased kernel size won't be accepted by embedded system, the possible i-cache usage/contention increase. And I've only done limited benchmark test, so I thought it may be safer to be off by default. Though my bold thought was it could be default on :) > It makes sense, but is it actually known that this does reduce the > variability? Yes, at lease for the strange performance bumps reported by 0day, like in [1][2][3]. > > as tracking these strange bump > > is quite time consuming. Though it can't help in other cases like data > > alignment changes like [4]. > > > > Following is some size data for v5.7 kernel built with a RHEL config > > used in 0day: > > > > text data bss dec filename > > 19738771 13292906 5554236 38585913 vmlinux.noalign > > 19758591 13297002 5529660 38585253 vmlinux.align32 > > > > Raw vmlinux size in bytes: > > > > v5.7 v5.7+align32 > > 253950832 254018000 +0.02% > > > > Some benchmark data, most of them have no big change: > > > > * hackbench: [ -1.8%, +0.5%] > > > > * fsmark: [ -3.2%, +3.4%] # ext4/xfs/btrfs > > > > * kbuild: [ -2.0%, +0.9%] > > > > * will-it-scale: [ -0.5%, +1.8%] # mmap1/pagefault3 > > > > * netperf: > > - TCP_CRR [+16.6%, +97.4%] > > - TCP_RR [-18.5%, -1.8%] > > - TCP_STREAM [ -1.1%, +1.9%] > > What do the numbers in [] mean? The TCP_CRR results look remarkable? For each of the benchmark listed above, I took 2 or 3 test platforms and run it with different parameters. So each of the benchmark will have several cases run, and [] lists the lowest and highest result. For the netperf/TCP_CRR case, the lowest is +16.6% on a Skylake server with 16 testing threads, and highest is +97.4 on a Cascadelake server with 96 testing threads. Thanks, Feng > > [1] https://lore.kernel.org/lkml/20200114085637.GA29297@shao2-debian/ > > [2] https://lore.kernel.org/lkml/20200330011254.GA14393@feng-iot/ > > [3] https://lore.kernel.org/lkml/1d98d1f0-fe84-6df7-f5bd-f4cb2cdb7f45@intel.com/ > > [4] https://lore.kernel.org/lkml/20200205123216.GO12867@shao2-debian/ > >
On Wed, Jul 22, 2020 at 08:39:19PM -0700, Andrew Morton wrote: > On Thu, 23 Jul 2020 11:30:01 +0800 Feng Tang <feng.tang@intel.com> wrote: > > > Recently 0day reported many strange performance changes (regression > > or improvement), in which there was no obvious relation between > > the culprit commit and the benchmark at the first look, and it causes > > people to doubt the test itself is wrong. > > > > Upon further check, many of these cases are caused by the change > > to the alignment of kernel text or data, as whole text/data of kernel > > are linked together, change in one domain may affect alignments of > > other domains. > > > > gcc has an option '-falign-functions=n' to force text aligned, and with > > that option enabled, some of those performance changes will be gone, > > like [1][2][3]. > > > > Add this option so that developers and 0day can easily find performance > > bump caused by text alignment change, > > Would they use it this way, or would they simply always enable the > option to reduce the variability? I may mis-understood it in my last reply. If you are asking about how will developers and 0day use this option, for 0day, I've talked with 0day folks, they may just enable it by default, as 0day cares more about the performance delta caused by a commit (Adding Philip and Rong from 0day). Thanks, Feng > It makes sense, but is it actually known that this does reduce the > variability? > > > as tracking these strange bump > > is quite time consuming. Though it can't help in other cases like data > > alignment changes like [4]. > > > > Following is some size data for v5.7 kernel built with a RHEL config > > used in 0day: > > > > text data bss dec filename > > 19738771 13292906 5554236 38585913 vmlinux.noalign > > 19758591 13297002 5529660 38585253 vmlinux.align32 > > > > Raw vmlinux size in bytes: > > > > v5.7 v5.7+align32 > > 253950832 254018000 +0.02% > > > > Some benchmark data, most of them have no big change: > > > > * hackbench: [ -1.8%, +0.5%] > > > > * fsmark: [ -3.2%, +3.4%] # ext4/xfs/btrfs > > > > * kbuild: [ -2.0%, +0.9%] > > > > * will-it-scale: [ -0.5%, +1.8%] # mmap1/pagefault3 > > > > * netperf: > > - TCP_CRR [+16.6%, +97.4%] > > - TCP_RR [-18.5%, -1.8%] > > - TCP_STREAM [ -1.1%, +1.9%] > > What do the numbers in [] mean? The TCP_CRR results look remarkable? > > > [1] https://lore.kernel.org/lkml/20200114085637.GA29297@shao2-debian/ > > [2] https://lore.kernel.org/lkml/20200330011254.GA14393@feng-iot/ > > [3] https://lore.kernel.org/lkml/1d98d1f0-fe84-6df7-f5bd-f4cb2cdb7f45@intel.com/ > > [4] https://lore.kernel.org/lkml/20200205123216.GO12867@shao2-debian/ > >
On Thu, 23 Jul 2020 14:29:33 +0800 Feng Tang <feng.tang@intel.com> wrote: > > > gcc has an option '-falign-functions=n' to force text aligned, and with > > > that option enabled, some of those performance changes will be gone, > > > like [1][2][3]. > > > > > > Add this option so that developers and 0day can easily find performance > > > bump caused by text alignment change, > > > > Would they use it this way, or would they simply always enable the > > option to reduce the variability? > > I may mis-understood it in my last reply. If you are asking about how > will developers and 0day use this option, for 0day, I've talked with > 0day folks, they may just enable it by default, as 0day cares more about > the performance delta caused by a commit (Adding Philip and Rong from > 0day). OK, thanks, I suspected as much. The patch is so simple and probably-will-work, I guess we toss it in there and see. However it would be good if the 0day people could use it for a while and then provide some feedback on whether it is actually proving useful. If not, we get to remove some stuff.
On Thu, Jul 23, 2020 at 05:57:04PM -0700, Andrew Morton wrote: > On Thu, 23 Jul 2020 14:29:33 +0800 Feng Tang <feng.tang@intel.com> wrote: > > > > > gcc has an option '-falign-functions=n' to force text aligned, and with > > > > that option enabled, some of those performance changes will be gone, > > > > like [1][2][3]. > > > > > > > > Add this option so that developers and 0day can easily find performance > > > > bump caused by text alignment change, > > > > > > Would they use it this way, or would they simply always enable the > > > option to reduce the variability? > > > > I may mis-understood it in my last reply. If you are asking about how > > will developers and 0day use this option, for 0day, I've talked with > > 0day folks, they may just enable it by default, as 0day cares more about > > the performance delta caused by a commit (Adding Philip and Rong from > > 0day). > > OK, thanks, I suspected as much. > > The patch is so simple and probably-will-work, I guess we toss it in > there and see. Thanks! > However it would be good if the 0day people could use it for a while > and then provide some feedback on whether it is actually proving > useful. If not, we get to remove some stuff. Yes, 0day is a good user to try this. Thanks, Feng
diff --git a/Makefile b/Makefile index 249a51d25c63..a59105e6f573 100644 --- a/Makefile +++ b/Makefile @@ -886,6 +886,10 @@ KBUILD_CFLAGS += $(CC_FLAGS_SCS) export CC_FLAGS_SCS endif +ifdef CONFIG_DEBUG_FORCE_FUNCTION_ALIGN_32B +KBUILD_CFLAGS += -falign-functions=32 +endif + # arch Makefile may override CC so keep this after arch Makefile is included NOSTDINC_FLAGS += -nostdinc -isystem $(shell $(CC) -print-file-name=include) diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug index 9ad9210d70a1..c1d52c4f120f 100644 --- a/lib/Kconfig.debug +++ b/lib/Kconfig.debug @@ -365,6 +365,17 @@ config SECTION_MISMATCH_WARN_ONLY If unsure, say Y. +config DEBUG_FORCE_FUNCTION_ALIGN_32B + bool "Force all function address 32B aligned" if EXPERT + help + There are cases that a commit from one domain changes the function + address alignment of other domains, and cause magic performance + bump (regression or improvement). Enable this option will help to + verify if the bump is caused by function alignment changes, while + it will slightly increase the kernel size and affect icache usage. + + It is mainly for debug and performance tuning use. + # # Select this config option from the architecture Kconfig, if it # is preferred to always offer frame pointers as a config
Recently 0day reported many strange performance changes (regression or improvement), in which there was no obvious relation between the culprit commit and the benchmark at the first look, and it causes people to doubt the test itself is wrong. Upon further check, many of these cases are caused by the change to the alignment of kernel text or data, as whole text/data of kernel are linked together, change in one domain may affect alignments of other domains. gcc has an option '-falign-functions=n' to force text aligned, and with that option enabled, some of those performance changes will be gone, like [1][2][3]. Add this option so that developers and 0day can easily find performance bump caused by text alignment change, as tracking these strange bump is quite time consuming. Though it can't help in other cases like data alignment changes like [4]. Following is some size data for v5.7 kernel built with a RHEL config used in 0day: text data bss dec filename 19738771 13292906 5554236 38585913 vmlinux.noalign 19758591 13297002 5529660 38585253 vmlinux.align32 Raw vmlinux size in bytes: v5.7 v5.7+align32 253950832 254018000 +0.02% Some benchmark data, most of them have no big change: * hackbench: [ -1.8%, +0.5%] * fsmark: [ -3.2%, +3.4%] # ext4/xfs/btrfs * kbuild: [ -2.0%, +0.9%] * will-it-scale: [ -0.5%, +1.8%] # mmap1/pagefault3 * netperf: - TCP_CRR [+16.6%, +97.4%] - TCP_RR [-18.5%, -1.8%] - TCP_STREAM [ -1.1%, +1.9%] [1] https://lore.kernel.org/lkml/20200114085637.GA29297@shao2-debian/ [2] https://lore.kernel.org/lkml/20200330011254.GA14393@feng-iot/ [3] https://lore.kernel.org/lkml/1d98d1f0-fe84-6df7-f5bd-f4cb2cdb7f45@intel.com/ [4] https://lore.kernel.org/lkml/20200205123216.GO12867@shao2-debian/ Signed-off-by: Feng Tang <feng.tang@intel.com> --- Makefile | 4 ++++ lib/Kconfig.debug | 11 +++++++++++ 2 files changed, 15 insertions(+)