Re: [RFT][patch 17/18] sched: use jump labels to reduce overhead when bandwidth control is inactive

[Paul Turner <pjt@google.com>]

On Thu, Jul 21, 2011 at 6:17 PM, Jason Baron <jbaron@redhat.com> wrote:
> On Thu, Jul 21, 2011 at 05:57:31PM -0700, Paul Turner wrote:
>> On Thu, Jul 21, 2011 at 5:32 PM, Jason Baron <jbaron@redhat.com> wrote:
>> > rth@redhat.com
>> > Bcc:
>> > Subject: Re: [RFT][patch 17/18] sched: use jump labels to reduce overhead
>> >  when bandwidth control is inactive
>> > Reply-To:
>> > In-Reply-To: <20110721184758.403388616@google.com>
>> >
>> > On Thu, Jul 21, 2011 at 09:43:42AM -0700, Paul Turner wrote:
>> >> So I'm seeing some strange costs associated with jump_labels; while on paper
>> >> the branches and instructions retired improves (as expected) we're taking an
>> >> unexpected hit in IPC.
>> >>
>> >> [From the initial mail we have workloads:
>> >>   mkdir -p /cgroup/cpu/test
>> >>   echo $$ > /dev/cgroup/cpu/test (only cpu,cpuacct mounted)
>> >>   (W1) taskset -c 0 perf stat --repeat 50 -e instructions,cycles,branches bash -c "for ((i=0;i<5;i++)); do $(dirname $0)/pipe-test 20000; done"
>> >>   (W2)taskset -c 0 perf stat --repeat 50 -e instructions,cycles,branches bash -c "$(dirname $0)/pipe-test 100000;true"
>> >>   (W3)taskset -c 0 perf stat --repeat 50 -e instructions,cycles,branches bash -c "$(dirname $0)/pipe-test 100000;"
>> >> ]
>> >>
>> >> To make some of the figures more clear:
>> >>
>> >> Legend:
>> >> !BWC = tip + bwc, BWC compiled out
>> >> BWC = tip + bwc
>> >> BWC_JL = tip + bwc + jump label (this patch)
>> >>
>> >>
>> >> Now, comparing under W1 we see:
>> >> W1: BWC vs BWC_JL
>> >>                             instructions            cycles                  branches              elapsed
>> >> ---------------------------------------------------------------------------------------------------------------------
>> >> clovertown [BWC]            845934117               974222228               152715407             0.419014188 [baseline]
>> >> +unconstrained              857963815 (+1.42)      1007152750 (+3.38)       153140328 (+0.28)     0.433186926 (+3.38)  [rel]
>> >> +10000000000/1000:          876937753 (+2.55)      1033978705 (+5.65)       160038434 (+3.59)     0.443638365 (+5.66)  [rel]
>> >> +10000000000/1000000:       880276838 (+3.08)      1036176245 (+6.13)       160683878 (+4.15)     0.444577244 (+6.14)  [rel]
>> >>
>> >> barcelona [BWC]             820573353               748178486               148161233             0.342122850 [baseline]
>> >> +unconstrained              817011602 (-0.43)       759838181 (+1.56)       145951513 (-1.49)     0.347462571 (+1.56)  [rel]
>> >> +10000000000/1000:          830109086 (+0.26)       770451537 (+1.67)       151228902 (+1.08)     0.350824677 (+1.65)  [rel]
>> >> +10000000000/1000000:       830196206 (+0.30)       770704213 (+2.27)       151250413 (+1.12)     0.350962182 (+2.28)  [rel]
>> >>
>> >> westmere [BWC]              802533191               694415157               146071233             0.194428018 [baseline]
>> >> +unconstrained              799057936 (-0.43)       751384496 (+8.20)       143875513 (-1.50)     0.211182620 (+8.62)  [rel]
>> >> +10000000000/1000:          812033785 (+0.27)       761469084 (+8.51)       149134146 (+1.09)     0.212149229 (+8.28)  [rel]
>> >> +10000000000/1000000:       811912834 (+0.27)       757842988 (+7.45)       149113291 (+1.09)     0.211364804 (+7.30)  [rel]
>> >> e.g. Barcelona issues ~0.43% less instructions, for a total of 817011602, in
>> >> the unconstrained case with BWC.
>> >>
>> >>
>> >> Where "unconstrained, 10000000000/1000, 10000000000/10000" are the on
>> >> measurements for BWC_JL, with (%d) being the relative difference to their
>> >> BWC counterparts.
>> >>
>> >> W1: BWC vs BWC_JL is very similar.
>> >>       BWC vs BWC_JL
>> >> clovertown [BWC]            985732031              1283113452               175621212             1.375905653
>> >> +unconstrained              979242938 (-0.66)      1288971141 (+0.46)       172122546 (-1.99)     1.389795165 (+1.01)  [rel]
>> >> +10000000000/1000:          999886468 (+0.33)      1296597143 (+1.13)       180554004 (+1.62)     1.392576770 (+1.18)  [rel]
>> >> +10000000000/1000000:       999034223 (+0.11)      1293925500 (+0.57)       180413829 (+1.39)     1.391041338 (+0.94)  [rel]
>> >>
>> >> barcelona [BWC]             982139920              1078757792               175417574             1.069537049
>> >> +unconstrained              965443672 (-1.70)      1075377223 (-0.31)       170215844 (-2.97)     1.045595065 (-2.24)  [rel]
>> >> +10000000000/1000:          989104943 (+0.05)      1100836668 (+0.52)       178837754 (+1.22)     1.058730316 (-1.77)  [rel]
>> >> +10000000000/1000000:       987627489 (-0.32)      1095843758 (-0.17)       178567411 (+0.84)     1.056100899 (-2.28)  [rel]
>> >>
>> >> westmere [BWC]              918633403               896047900               166496917             0.754629182
>> >> +unconstrained              914740541 (-0.42)       903906801 (+0.88)       163652848 (-1.71)     0.758050332 (+0.45)  [rel]
>> >> +10000000000/1000:          927517377 (-0.41)       952579771 (+5.67)       170173060 (+0.75)     0.771193786 (+2.43)  [rel]
>> >> +10000000000/1000000:       914676985 (-0.89)       936106277 (+3.81)       167683288 (+0.22)     0.764973632 (+1.38)  [rel]
>> >>
>> >> Now this is rather odd, almost across the board we're seeing the expected
>> >> drops in instructions and branches, yet we appear to be paying a heavy IPC
>> >> price.  The fact that wall-time has scaled equivalently with cycles roughly
>> >> rules out the cycles counter being off.
>> >>
>
> if i understand your results, for barcelona you did see an improvement
> in cycles and eslapsed time with jump labels for unconstrained?
>

Under W2, yes.

>> >> We are seeing the expected behavior in the bandwidth enabled case;
>> >> specifically the <jl=jmp><ret><cond><ret> blocks are taking an extra branch
>> >> and instruction which shows up on all the numbers above.
>> >>
>> >> With respect to compiler mangling the text is essentially unchanged in size.
>> >> One lurking suspicion is whether the inserted nops have perturbed some of the
>> >> jmp/branch alignments?
>
> hmmmm....not sure, I'm adding Richard Henderson, to the 'cc list, who
> worked on the 'asm goto' in gcc.
>
>> >>
>> >>     text    data     bss     dec     hex filename
>> >>  7277206 2827256 2125824 12230286         ba9e8e vmlinux.jump_label
>> >>  7276886 2826744 2125824 12229454         ba9b4e vmlinux.no_jump_label
>> >>
>
> the other thing here is that vmlinux.jump_label includes the extra
> kernel/jump_label.o file, so you can sort of subtract the text size of
> that file to do a fair comparison.

Even without doing that it's only a 1.00004% change in text size.

I was just making the inference that if it's gcc mangling it's likely
in the layout/alignment.

>
> Also, I would have expected the data section to have increased more with
> jump labels enabled. Are tracepoints disabled (a current user of jump
> labels).

Yeah -- Tracing is enabled so the BWC build should have labels
already; this likely accounts for the small increase noted above.

>
>> >>  I have checked to make sure that the right instructions are being patched in
>> >>  at run-time.  I've also pulled a fully patched jump_label out of the kernel
>> >>  into a userspace test (and benchmarked it directly under perf).  The results
>> >>  here are also exactly as expected.
>> >>
>> >> e.g.
>> >>  Performance counter stats for './jump_test':
>> >>      1,500,839,002 instructions, 300,147,081 branches 702,468,404 cycles
>> >> Performance counter stats for './jump_test 1':
>> >>      2,001,014,609 instructions, 400,177,192 branches 901,758,219 cycles
>> >>
>
> what no-op did you use in userspace? I wouldn't think the no-op choice
> would make any difference though...At compile time we use a 'jmp 0', and
> then at boot we dynamically patch the 'jmp 0' with the no-op we think works
> best...
>

Sorry -- what I meant here is I pulled the run-time chosen "best" nop
out of /proc/kcore and tested a
tight loop about a <JL><RET><COND><RET> sequence (e.g.
cfs_rq_throttled()) with JL being the nop and jmp respectively.

Specifically for Westmere this ends up being K8_NOP5  -- 0x666666D0

> thanks,
>
> -Jason
>
>> >> Overall if we can fix the IPC the benefit in the globally unconstrained case
>> >> looks really good.
>> >>
>> >> Any thoughts Jason?
>> >>
>> >
>> > Do you have CONFIG_CC_OPTIMIZE_FOR_SIZE set? I know that when
>> > CONFIG_CC_OPTIMIZE_FOR_SIZE is not set, the compiler can make the code
>> > more optimal.
>> >
>>
>> Ah I should have mentioned that was one of the holes I stared down:
>>
>> Builds were -O2 (gcc-4.6.1) and
>> $  zcat /proc/config.gz | grep CONFIG_CC_OPTIMIZE_FOR_SIZE
>> # CONFIG_CC_OPTIMIZE_FOR_SIZE is not set
>>
>> Same kernel image across all platforms.
>>
>>
>>
>>
>>
>>
>> > thanks,
>> >
>> > -Jason
>> >
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>> Please read the FAQ at  http://www.tux.org/lkml/
>