lkml.org 
[lkml]   [2001]   [Dec]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: IO degradation in 2.4.17-pre2 vs. 2.4.16
Sure, I wasn't protesting the patch or anything; I was just passing my
observations along. I also couldn't care less about dbench numbers for
the sake of dbench numbers; I was just using it and other simple
benchmarks as stepping stones to try to figure out what effect bdflush
and max_readahead settings actually have on the way the system
performs. After the simple benchmarks narrowed things down, I would've
run more exhaustive benchmarks, then some large MM5 runs (including
setup, takedown, post-processing into graphs, etc), enough Gaussian jobs
to create 200 GB or so of scratch files, a hundred or so BLAST jobs
against a centralized database, all or part of these at the same time,
etc, the typical stuff that I see running. If I were to start out with
the real workload it'd take years.

The thing is, everywhere I read about tweaking filesystem performance
someone has some magic number to throw into bdflush. There's never any
justification for it and it's 9 times out of 10 for a "server" system,
whatever that is. Some recommendations are for values larger than
fs/buffer.c allows, some are wacko recommending 100/100 for
nfract/nfract_sync, some want 5000 or 6000 for nfract_sync, which seems
somehow wrong for a percentage (perhaps older kernels didn't have a
percentage there or something). There are even different bdflush
numbers between 2.4.13-pre2, 2.4.17, and 2.4.17-pre1aa1. I was just
looking for a way to profile the way the different settings affect
system performance under a variety of conditions and dbench seemed like
a way to get the 'many clients / many small files' aspect of it all.
Who knows, maybe the default numbers are the best compromise or maybe
the continuing vm tweaks will make any results from a previous kernel
invalid for a current kernel or maybe the bdflush tweaking isn't really
worth it at all and I'm better off getting on with mucking about with
larger hardware and parallel filesystems. At least I learned that I
really do want a larger max_readahead number.

As for interactivity, if the changes have any effect on the number of
"NFS server blah not responding" messages I get, I'll be more than
happy.

Thanks,

--
Jason Holmes

Marcelo Tosatti wrote:
>
> Jason,
>
> Yes, throughtput-only tests will have their numbers degradated with the
> change applied on 2.4.16-pre2.
>
> The whole thing is just about tradeoffs: Interactivity vs throughtput.
>
> I'm not going to destroy interactivity for end users to get beatiful
> dbench numbers.
>
> And about your clients: Don't you think they want some kind of
> decent latency on their side?
>
> Anyway, thanks for your report!
>
> On Sat, 1 Dec 2001, Jason Holmes wrote:
>
> > I saw in a previous thread that the interactivity improvements in
> > 2.4.17-pre2 had some adverse effect on IO throughput and since I was
> > already evaluating 2.4.16 for a somewhat large fileserving project, I
> > threw 2.4.17-pre2 on to see what has changed. Throughput while serving
> > a large number of clients is important to me, so my tests have included
> > using dbench to try to see how things scale as clients increase.
> >
> > 2.4.16:
> >
> > Throughput 116.098 MB/sec (NB=145.123 MB/sec 1160.98 MBit/sec) 1 procs
> > Throughput 206.604 MB/sec (NB=258.255 MB/sec 2066.04 MBit/sec) 2 procs
> > Throughput 210.364 MB/sec (NB=262.955 MB/sec 2103.64 MBit/sec) 4 procs
> > Throughput 213.397 MB/sec (NB=266.747 MB/sec 2133.97 MBit/sec) 8 procs
> > Throughput 210.989 MB/sec (NB=263.736 MB/sec 2109.89 MBit/sec) 16
> > procs
> > Throughput 138.713 MB/sec (NB=173.391 MB/sec 1387.13 MBit/sec) 32
> > procs
> > Throughput 117.729 MB/sec (NB=147.162 MB/sec 1177.29 MBit/sec) 64
> > procs
> > Throughput 66.7354 MB/sec (NB=83.4193 MB/sec 667.354 MBit/sec) 128
> > procs
> >
> > 2.4.17-pre2:
> >
> > Throughput 96.2302 MB/sec (NB=120.288 MB/sec 962.302 MBit/sec) 1 procs
> > Throughput 226.679 MB/sec (NB=283.349 MB/sec 2266.79 MBit/sec) 2 procs
> > Throughput 223.955 MB/sec (NB=279.944 MB/sec 2239.55 MBit/sec) 4 procs
> > Throughput 224.533 MB/sec (NB=280.666 MB/sec 2245.33 MBit/sec) 8 procs
> > Throughput 153.672 MB/sec (NB=192.09 MB/sec 1536.72 MBit/sec) 16 procs
> > Throughput 91.3464 MB/sec (NB=114.183 MB/sec 913.464 MBit/sec) 32
> > procs
> > Throughput 64.876 MB/sec (NB=81.095 MB/sec 648.76 MBit/sec) 64 procs
> > Throughput 54.9774 MB/sec (NB=68.7217 MB/sec 549.774 MBit/sec) 128
> > procs
> >
> > Throughput 136.522 MB/sec (NB=170.652 MB/sec 1365.22 MBit/sec) 1 procs
> > Throughput 223.682 MB/sec (NB=279.603 MB/sec 2236.82 MBit/sec) 2 procs
> > Throughput 222.806 MB/sec (NB=278.507 MB/sec 2228.06 MBit/sec) 4 procs
> > Throughput 224.427 MB/sec (NB=280.534 MB/sec 2244.27 MBit/sec) 8 procs
> > Throughput 152.286 MB/sec (NB=190.358 MB/sec 1522.86 MBit/sec) 16
> > procs
> > Throughput 92.044 MB/sec (NB=115.055 MB/sec 920.44 MBit/sec) 32 procs
> > Throughput 78.0881 MB/sec (NB=97.6101 MB/sec 780.881 MBit/sec) 64
> > procs
> > Throughput 66.1573 MB/sec (NB=82.6966 MB/sec 661.573 MBit/sec) 128
> > procs
> >
> > Throughput 117.95 MB/sec (NB=147.438 MB/sec 1179.5 MBit/sec) 1 procs
> > Throughput 212.469 MB/sec (NB=265.586 MB/sec 2124.69 MBit/sec) 2 procs
> > Throughput 214.763 MB/sec (NB=268.453 MB/sec 2147.63 MBit/sec) 4 procs
> > Throughput 214.007 MB/sec (NB=267.509 MB/sec 2140.07 MBit/sec) 8 procs
> > Throughput 96.6572 MB/sec (NB=120.821 MB/sec 966.572 MBit/sec) 16
> > procs
> > Throughput 48.1342 MB/sec (NB=60.1677 MB/sec 481.342 MBit/sec) 32
> > procs
> > Throughput 71.3444 MB/sec (NB=89.1806 MB/sec 713.444 MBit/sec) 64
> > procs
> > Throughput 59.258 MB/sec (NB=74.0724 MB/sec 592.58 MBit/sec) 128 procs
> >
> > I have included three runs for 2.4.17-pre2 to show how inconsistent its
> > results are; 2.4.16 didn't have this problem to this extent. bonnie++
> > numbers seem largely unchanged between kernels, coming in around:
> >
> > ------Sequential Output------ --Sequential Input- --Random-
> > -Per Chr- --Block-- -Rewrite- -Per Chr- --Block-- --Seeks--
> > Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP /sec %CP
> > 2512M 14348 81 49495 26 24438 16 16040 96 55006 15 373.7 1
> > ------Sequential Create------ --------Random Create--------
> > -Create-- --Read--- -Delete-- -Create-- --Read--- -Delete--
> > files /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP
> > 16 3087 99 +++++ +++ +++++ +++ 3175 100 +++++ +++ 11042 100
> >
> > The test machine is an IBM 342 with 2 1.26 GHz P3 processors and 1.25 GB
> > of RAM. The above numbers were generated off of 1 10K RPM SCSI disk
> > hanging off of an Adaptec aix7899 controller.
> >
> > --
> > Jason Holmes
> > -
> > 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/
> >
-
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/

\
 
 \ /
  Last update: 2005-03-22 13:18    [W:0.112 / U:6.308 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site