[Fwd: Re: Testing of Reiserfs write speedup patch]

Don't know if this is useful or not... I thought I would send my testing results to you guys. I have tested the heck out of my system after applying the ReiserFS_speedup-2.4.19-pre6.diff patch to AA's 2.4.19-rc5-aa1. This is actually aa's version of the released 2.4.19, since 2.4.19-rc5 is the 2.4.19 release. Of note is that AA's kernel includes aio and the O(1) scheduler.

I have not noted any fs corruption or other problems during the testing.

System Specs relevant:

Tyan S2460 Tiger MP motherboard
2xAthlon MP 1600+ processors
1.5 GB RAM
Mylex AcceleRAID 170 RAID controller with 32 MB onboard cache, running 5 IBM DDYS-T18350N 10k rpm U160 SCSI drives in a RAID 5 configuration.

Note /proc/sys/vm/max-readahead in all cases set to 511.

First bonnie++ (2.4.19-aa1 w/o speedup patch)
Version 1.01d ------Sequential Output------ --Sequential Input-
--Random-

                   -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
jco-linux 3100M 21468 33 28510 20 17655 10 38814 65 82601 40
399.0 3

                   ------Sequential Create------ --------Random
Create--------
                   -Create-- --Read--- -Delete-- -Create-- --Read---
-Delete--
             files  /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP
/sec %CP
                16 17419  56 +++++ +++ 16464  58 17617  59 +++++ +++

7220 40 jco-linux,3100M,21468,33,28510,20,17655,10,38814,65,82601,40,399.0,3,16,17419,56,+++++,+++,16464,58,17617,59,+++++,+++,7220,40

Second bonnie++ (2.4.19-aa1 w/ speedup patch)
Version 1.01d ------Sequential Output------ --Sequential Input-
--Random-

                   -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
jco-linux 3100M 37048 58 39513 30 18319 11 40981 68 93992 47
373.0 3

                   ------Sequential Create------ --------Random
Create--------
                   -Create-- --Read--- -Delete-- -Create-- --Read---
-Delete--
             files  /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP
/sec %CP
                16 17602  90 +++++ +++ 15377  92 14751  83 +++++ +++

13464 94 jco-linux,3100M,37048,58,39513,30,18319,11,40981,68,93992,47,373.0,3,16,17602,90,+++++,+++,15377,92,14751,83,+++++,+++,13464,94

Notice the tremendous speedup on writes, and even a decent speedup on sequential reads! I'm a happy man here! (94 MB/sec reads are real rates from the combined output of the drives, since I am using a file size more than double the memory size. The peak transfer I see during actual physical media sequential reads from the controller is about 105 MB/sec, pretty close to saturation for the 32-bit PCI bus.) I think the combined real sequential output of four DDYS's (i.e. not including the parity equivalent) is a bit higher than 100 MB/sec, but the PCI bus is constraining here. (Too bad I don't have the AcceleRAID 352 or ExtremeRAID 64 bit controllers.)

During bonnie++ runs without the speedup patch, the system is really prone to read starvation during the big writes. The behaviour with the speed-up patch is vastly improved (although not perfect).

Dbench results with speedup patch applied (without any tweaking to the /proc/sys/vm settings):

Throughput 68.9744 MB/sec (NB=86.2179 MB/sec  689.744 MBit/sec)  8 procs
Throughput 65.0722 MB/sec (NB=81.3403 MB/sec  650.722 MBit/sec)  16 procs
Throughput 43.4186 MB/sec (NB=54.2732 MB/sec  434.186 MBit/sec)  32 procs
Throughput 35.9702 MB/sec (NB=44.9627 MB/sec  359.702 MBit/sec)  64 procs

I ran dbenchs all the way up to 256. Throughput asymptoted out at ~30 MB/sec. These results are consistent with runs I did a while back with the 2.4.18 kernel. No problems were observed.

I have also run the AIM suite7 benchmark, since it has now been released as open source. The results are attached as a text and ps file. A result of peak 1059.7 jobs/minute, with a sustained rate of 958.2 jobs/minute seems excellent to me. The machine was still running at a total of 930.2 jobs/minute at the crossover point, where 963 simultaneous multiuser processes were running. (Yep, the load was at ~1000.) Believe it or not, the system GUI was still very usable, thanks to the O(1) scheduler.

I ran the same benchmark on a dual PIII 700 with the SuSE 2.4.18 default kernel and a single Maxtor 30 GB hard drive, and the peak rate was at 135 jobs/minute, a factor of 7.85 difference from the dual Athlon RAID system with the patches.

Anyway, I was more interested in testing stability than speed on the dual Athlon. The AIM benchmark completed with no problems whatsover. Not a single error or warning logged in /var/log/messages. File system is perfectly fine - no corruption or other problems.

As a result of this testing, looks like the write speedup patch is a keeper, at least for me!

Jim Owens

------------------------------------------------------------------------

Tasks Jobs/Min JTI Real CPU Jobs/sec/task
Benchmark Version Machine Run Date
AIM Multiuser Benchmark - Suite VII "1.1" JCO-LINUX Sep 23 10:10:36 2002
1 101.4 100 57.4 7.6 1.6893
2 195.2 99 59.6 13.3 1.6267
3 301.3 99 57.9 16.4 1.6741
4 346.7 98 67.1 17.3 1.4447
5 416.1 97 69.9 24.3 1.3871
6 484.3 97 72.1 24.6 1.3454
7 535.7 97 76.0 32.7 1.2755
8 594.5 98 78.3 36.5 1.2385
10 675.6 97 86.2 43.1 1.1259
14 761.6 97 107.0 59.7 0.9067
18 835.3 97 125.4 74.6 0.7735
26 919.0 97 164.7 107.2 0.5891
29 944.6 96 178.7 111.8 0.5429
35 995.9 96 204.5 137.9 0.4742
47 1031.3 96 265.2 182.1 0.3657
59 1030.2 94 333.3 232.7 0.2910
71 1055.6 94 391.5 274.1 0.2478
97 1046.9 94 539.2 374.3 0.1799
107 1014.1 95 614.1 413.4 0.1580
128 973.3 96 765.4 487.4 0.1267
149 1031.8 96 840.5 575.4 0.1154
193 1017.3 95 1104.2 739.1 0.0878
212 1059.7 95 1164.3 816.5 0.0833
251 1023.3 92 1427.6 963.2 0.0679
290 994.8 93 1696.6 1111.3 0.0572
375 1003.2 93 2175.5 1431.0 0.0446
460 981.1 94 2728.8 1739.8 0.0355
545 995.1 92 3187.5 2082.8 0.0304
729 956.0 92 4438.0 2792.4 0.0219
805 968.8 92 4835.9 3099.0 0.0201
963 930.2 92 6025.0 3718.4 0.0161