Skip to contentSkip to conversation

Linux 2.4.18-rc3-jam1

PATCH [email protected] 2 participants 2 messages Feb 24, 2002

lore
Share & tools
lore
#0 lore raw
Quoted 9 lines
Quoted 7 lines

Random Writes

                  Num                 Avg      Maximum     Lat%    Lat%  CPU
Kernel            Thr Rate  (CPU%)  Latency    Latency      >2s    >10s  Eff
----------------- --- ------------------------------------------------------
2.4.18-rc2        128  0.67  1.87%    1.334     777.23  0.00000  0.00000  36
2.4.18-rc2-jam1   128  0.80  5.72%    0.190       3.68  0.00000  0.00000  14
2.4.18rc2aa2      128  0.61  1.39%   61.796   72674.58  0.32761  0.32761  44

Holy cow! Are you sure these numbers are right?

Max latency for random writes at 128 threads varies widely. The numbers above happen to hit on the widest variation.

Random Writes

                  Num                 Avg      Maximum     Lat%    Lat%  CPU
Kernel            Thr Rate  (CPU%)  Latency    Latency      >2s    >10s  Eff
----------------- --- ------------------------------------------------------
2.4.17-rmap12e    128  0.71  5.19%    1.689     981.25  0.00000  0.00000  14
2.4.17-rmap12f    128  0.71  4.97%    0.975     917.19  0.00000  0.00000  14
2.4.18-pre9       128  0.70  1.82%    1.664     785.13  0.00000  0.00000  38
2.4.18-pre9-ac1   128  0.67  1.93%    0.339     398.63  0.00000  0.00000  35
2.4.18-pre9-ac3   128  0.71  5.09%    0.438     951.04  0.00000  0.00000  14
2.4.18-pre9-ac4   128  0.72  5.12%    0.191      11.63  0.00000  0.00000  14
2.4.18-pre9-am1   128  0.76  2.72%    1.527     845.77  0.00000  0.00000  28
2.4.18-pre9-am2   128  0.75  2.83%    1.371     886.19  0.00000  0.00000  27
2.4.18-pre9-mjc2  128  0.71  3.72%    0.220      12.10  0.00000  0.00000  19
2.4.18-rc1        128  0.68  1.81%    1.584     800.21  0.00000  0.00000  37
2.4.18-rc1-slb1   128  0.81  5.77%    0.192       5.25  0.00000  0.00000  14
2.4.18-rc2        128  0.67  1.87%    1.334     777.23  0.00000  0.00000  36
2.4.18-rc2-ac2    128  0.71  5.19%    0.340     555.11  0.00000  0.00000  14
2.4.18-rc2-jam1   128  0.80  5.72%    0.190       3.68  0.00000  0.00000  14
2.4.18-rc4-jam1   128  0.80  5.72%    5.025    6734.19  0.07560  0.00000  14
2.4.18rc2aa2      128  0.61  1.39%   61.796   72674.58  0.32761  0.32761  44
2.5.3-dj5-tio     128  0.75  2.17%    0.209       1.73  0.00000  0.00000  34
2.5.4-dj2         128  0.76  2.23%    0.206       1.83  0.00000  0.00000  34
2.5.4-dj3         128  0.75  2.30%    0.207       1.87  0.00000  0.00000  33
2.5.5             128  0.74  2.09%    0.197       1.97  0.00000  0.00000  35
2.5.5-dj1         128  0.76  2.26%    0.201       1.45  0.00000  0.00000  33
2.5.5-pre1        128  0.75  2.28%    0.197       2.64  0.00000  0.00000  33
Quoted 8 lines

The increased throughput will be thanks to the boosted request queue size.

The (greatly) increased CPU load will also be due to browsing the eight-times larger request queue. Plus we browse it a bit more than we used to.

The improvement in worst-case latency in both -aa and -jam will be due to the FIFO wait for requests.

Thanks for the insight there.

Quoted 5 lines

But improvement by a factor of 20,000 sounds a little excessive :) And a maximum latency of three milliseconds would seem to indicate that the benchmark is *never* waiting on disk seek, and that perhaps the request queue is simply never filling up. But that doesn't make sense.

I notice the 2.5.x kernels also have extremely low max latency on this test.

Quoted 2 lines

What does the "latency" actually mean? Is it the time spent in the kernel to issue a write(2)?

AFAICT, it's the time between when a request like write(2) is made, and when it completes. It doesn't appear to be actual time in kernel. I'd rather hear from a more knowledgeable coder though.

Quoted 2 lines

Something funny is happening, I suspect. Guess I should go look at tiobench...

Cool. I admire your work.

Below is a snippet of tiobench on random writes. The rc4-jam1 included the entire patchset, whereas rc2-jam1 had patches with the first two digits < 20.

Most notable is that jam1 had an unusual low max latency at 16 threads, which didn't appear at 8, 32, 64, and 128.

Random Writes

                 Num                  Avg      Maximum     Lat%     Lat% CPU
Kernel           Thr  Rate  (CPU%)  Latency    Latency      >2s     >10s Eff
--------------------  ------------------------------------------------------
2.4.18-rc2-jam1    8   0.71  4.71%    0.190       3.53  0.00000  0.00000  15
2.4.18-rc4-jam1    8   0.71  4.61%    0.754    2242.97  0.02500  0.00000  15
2.4.18-rc2-jam1   16   0.72  4.76%    0.192       4.02  0.00000  0.00000  15
2.4.18-rc4-jam1   16   0.72  4.72%    0.196       5.09  0.00000  0.00000  15
2.4.18-rc2-jam1   32   0.77  5.13%    0.192       5.28  0.00000  0.00000  15
2.4.18-rc4-jam1   32   0.77  5.25%    8.268   13672.38  0.15000  0.00000  15
2.4.18-rc2-jam1   64   0.79  5.27%    0.190       3.63  0.00000  0.00000  15
2.4.18-rc4-jam1   64   0.78  5.30%   26.660   27314.64  0.50403  0.02520  15
2.4.18-rc2-jam1  128   0.80  5.72%    0.190       3.68  0.00000  0.00000  14
2.4.18-rc4-jam1  128   0.80  5.72%    5.025    6734.19  0.07560  0.00000  14
Signature · 1 line
Randy Hron
#1 lore raw

Hi.

First of all, I never expected this tniny patch collection would origin such "rivers of e-ink"...
Please, have always present that I can have made some mistake in my offset re-engineering. But the fact that it at least survives the tests is good.

Quoted 2 lines · On 20020224 [email protected] wrote:
2.4.18-rc2-jam1   128  0.80  5.72%    0.190       3.68  0.00000  0.00000  14
2.4.18-rc4-jam1   128  0.80  5.72%    5.025    6734.19  0.07560  0.00000  14

This is really strange. I have looked at my patches and are the same. What changed in mainlaine ??

[...]

Quoted 5 lines

Below is a snippet of tiobench on random writes. The rc4-jam1 included the entire patchset, whereas rc2-jam1 had patches with the first two digits < 20.

So rc2-jam1 is running without the ide-update (I noticed your system is IDE), but also without irqrate.
I will reorder the patches so you can apply 0*, 1* and 3* without problems, then scsi-ide updates. And you can try with/out the ide update. Do not see what can be related with latency in >=20*, apart from ide and irqrate...

I do not know if you are already doing this, but I will skip the bproc part for thins tests. It pollutes system calls with hooks for network, so it can be hurting in many ways.

I will release a -jam2 with latest vm-27 and reordering the patches.

Signature · 4 lines
J.A. Magallon                           #  Let the source be with you...        
mailto:[email protected]
Mandrake Linux release 8.2 (Cooker) for i586
Linux werewolf 2.4.18-rc4-jam1 #1 SMP Sat Feb 23 16:25:56 CET 2002 i686