Impact of HIGHMEM?

Impact of HIGHMEM?

[war]

Does anyone have any benchmarks as to how much HIGHMEM affects
performance in Linux?

Searched google.com + groups.google.com, couldn't find anything solid
though.

Re: Impact of HIGHMEM?

[Ivanovich]

A Dissabte 08 Desembre 2001 20:03, war va escriure:
> Does anyone have any benchmarks as to how much HIGHMEM affects
> performance in Linux?
>
> Searched google.com + groups.google.com, couldn't find anything solid
> though.

why don't you try to compile a kernel with HIGHMEM and another without it and 
then run some benchmarks in each one and compare?

not everyone have the amount of ram to test this (i only have 
256...(sigh)) if i had that amount i would run some bench...

Re: Impact of HIGHMEM?

[Jens Axboe]

On Sat, Dec 08 2001, Ivanovich wrote:
> A Dissabte 08 Desembre 2001 20:03, war va escriure:
> > Does anyone have any benchmarks as to how much HIGHMEM affects
> > performance in Linux?
> >
> > Searched google.com + groups.google.com, couldn't find anything solid
> > though.
> 
> why don't you try to compile a kernel with HIGHMEM and another without it and 
> then run some benchmarks in each one and compare?

You'll very quickly spend a significant amount of sys time copying pages
back and forth. 

> not everyone have the amount of ram to test this (i only have 
> 256...(sigh)) if i had that amount i would run some bench...

You don't need lots of mem to test highmem impact, just grab the highmem
debug patch from Andrea:

kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.17pre4aa1/20_highmem-debug-7

-- 
Jens Axboe

Re: Impact of HIGHMEM?

[Ivanovich]

A Dissabte 08 Desembre 2001 21:05, Jens Axboe va escriure:
> On Sat, Dec 08 2001, Ivanovich wrote:
> > A Dissabte 08 Desembre 2001 20:03, war va escriure:
> > > Does anyone have any benchmarks as to how much HIGHMEM affects
> > > performance in Linux?
> You don't need lots of mem to test highmem impact, just grab the highmem
> debug patch from Andrea:
>
> kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.17pre4aa1/20_hig
>hmem-debug-7

ok, i made my homeworks, have tested highmem effect in my computer but i get 
some STRANGE results

kernel: 2.4.17-pre6 + preempt-kernel-rml-2.4.17-pre6-1.patch + 
lock-break-rml-2.4.17-pre6-1.patch (+ highmem for lowmem machines patch)
micro: pIII @ 935 Mhz
mem: 256Mb
L2 cache: 256Kb

i run cachebench (llcbench suite) Read/Modify/Write benchmark

results WITHOUT highmem:
size - Mb/sec
32768 - 4290.175472
49152 - 4277.256065
65536 - 4293.230743
98304 - 4294.457415
131072 - 4291.695324 <-
196608 - 2327.512263 <- 
262144 - 1177.041779 <-
393216 - 709.747398 <-
524288 - 591.464925 <-
786432 - 558.824848
1048576 - 573.862697
1572864 - 584.174573
2097152 - 575.021368
3145728 - 580.856111
4194304 - 581.096127
6291456 - 587.200702
8388608 - 583.293192
12582912 - 583.264028
16777216 584.326629

results WITH highmem:
size - Mb/sec
32768 - 4290.283522 <- all small values 
49152 - 4277.199336 <- similar to no-highmem
65536 - 4293.282408 <- due to cache
98304 - 4294.589967
131072 - 4291.831538 <- strange...
196608 - 4295.502217 <- 
262144 - 2020.226275 <- 
393216 - 605.011875 <-
524288 - 454.203301 <-
786432 - 438.363738
1048576 - 438.401140
1572864 - 438.595614
2097152 - 438.794817
3145728 - 439.164867
4194304 - 439.566675
6291456 - 440.364513
8388608 - 441.159194
12582912 - 442.724326
16777216 - 444.301898

aprox. 24% throughput drop with highmem in the read/modify/write bench when 
cache have no effect. of course this is a synthetic

but quite strange (for me) behaviour in the lines marked with "???":
no-highmem
131072 - 4291.695324 <-
196608 - 2327.512263 <- ???
262144 - 1177.041779 <- ???
393216 - 709.747398 <-
524288 - 591.464925 <-
highmem
131072 - 4291.831538 <-
196608 - 4295.502217 <- ???
262144 - 2020.226275 <- ???
393216 - 605.011875 <- 
524288 - 454.203301 <-

can anyone explain me why this happen? himem seems to be faster sometimes, 
and cache have something to do.... i have 256Kb L2