linux-kernel - IO CPU Affinity: more results...

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [day] [month] [year] [list]

Message-ID: <47EA94C9.4060308@hp.com>
Date:	Wed, 26 Mar 2008 14:24:09 -0400
From:	"Alan D. Brunelle" <Alan.Brunelle@...com>
To:	linux-kernel@...r.kernel.org
Cc:	Jens Axboe <jens.axboe@...cle.com>, npiggin@...e.de, dgc@....com
Subject: IO CPU Affinity: more results...

Current blk.git origin/io-cpu-affinity sources:

After 60 successful passes on a 4-way ia64:

 60 0 mkfs untar make 
 60 1 mkfs untar make 

   Part RQ     Min     Avg     Max     Dev
  ----- -- ------- ------- ------- -------
   mkfs  0  81.233  81.810  82.599   0.330 
   mkfs  1  81.083  81.854  82.973   0.405

  untar  0  17.075  17.676  18.098   0.273 
  untar  1  16.975  17.570  18.128   0.288

   make  0  24.231  24.380  24.541   0.085 
   make  1  24.116  24.312  24.459   0.073 
  ----- -- ------- ------- ------- -------
   comb  0 122.898 123.866 125.002   0.516 
   comb  1 122.620 123.736 125.156   0.521 
  ===== == ======= ======= ======= =======
   psys  0   2.12%   2.19%   2.28%   0.035 
   psys  1   1.92%   2.00%   2.25%   0.049 

so:

1. It's working pretty solidly on ia64
2. We still see reduced combined times w/ rq=1 (albeit, not much - and certainly nothing definitive with the relatively high standard deviations).
3. We see large reduction in %system to do the same work - 8% less

And here's something else I've noticed: It seems that as the runs go on, the makes happen quicker (in general). I've got some graphs on free.linux.hp.com - they are a bit busy, but here are some pointers:

1. Black stuff is for rq=0, red stuff is for rq=1
2. Solid horizontal line indicates the set average.
3. Lower numbers for /all/ graphs are better
4. Open circles represent individual test run points
5. Hashed-line w/ shaded large circles represents localized averages: each circle is the average of the surrounding 5 data points.

The last thing is key: you'll see on 

http://free.linux.hp.com/~adb/jens/make.png

The black hashed line (rq=0) tends to bop around the average, whilst the red hashed line (rq=1) seems to show a downwards trend. (I need to run this a lot longer to see if it holds up.) Note: this trending-downwards does /not/ appear in the mkfs & untar parts of the operations. /Every/ time I've had extended runs with rq=1 I see this trend. 

Note: %sys doesn't really fluctuate much at all - as can be seen by: 

http://free.linux.hp.com/~adb/jens/psys.png

The other graphs include:

http://free.linux.hp.com/~adb/jens/mkfs.png
http://free.linux.hp.com/~adb/jens/untar.png
http://free.linux.hp.com/~adb/jens/comb.png

Alan D. Brunelle
HP / OSLO / S&P

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/