dd test discussion (was Re: possible regression?)
On Fri, Jan 21, 2011 at 20:25, Mag Gam <magawake@gmail.com> wrote:
Typo, in Box 5. Wrong kernel version
Box 5, RHEL 5.4, 2.6.36, dd bs=1024 count=1000000 if=/dev/zero bs=4096k of=/dev/null
Anyone had a chance to try this? :-)
$ for a in `seq 1 3`; do \time dd bs=4K count=1M if=/dev/zero of=/dev/null; done 1048576+0 records in 1048576+0 records out 4294967296 bytes (4.3 GB) copied, 1.38893 seconds, 3.1 GB/s 0.21user 1.17system 0:01.39elapsed 99%CPU (0avgtext+0avgdata 2768maxresident)k 0inputs+0outputs (0major+209minor)pagefaults 0swaps 1048576+0 records in 1048576+0 records out 4294967296 bytes (4.3 GB) copied, 1.39358 seconds, 3.1 GB/s 0.23user 1.16system 0:01.39elapsed 99%CPU (0avgtext+0avgdata 2768maxresident)k 0inputs+0outputs (0major+209minor)pagefaults 0swaps 1048576+0 records in 1048576+0 records out 4294967296 bytes (4.3 GB) copied, 1.4527 seconds, 3.0 GB/s 0.17user 1.23system 0:01.45elapsed 96%CPU (0avgtext+0avgdata 2752maxresident)k 0inputs+0outputs (0major+208minor)pagefaults 0swaps I slightly modified your dd test. Again, this is on 2.6.36. IMO, you shouldn't worry much about it. Were experiencing something that related to memory access? It's not a NUMA box, right? Or are you using PAE? just capping into 896 MB maybe? non standart VM split other than 3:1? PS: I rewrite the subject to better match the discussion. With the hope readers well get better understanding about what the discussion is all about. Hope you don't mind.... -- regards, Mulyadi Santosa Freelance Linux trainer and consultant blog: the-hydra.blogspot.com training: mulyaditraining.blogspot.com
On Sat, Jan 22, 2011 at 11:49:43AM +0700, Mulyadi Santosa wrote:
On Fri, Jan 21, 2011 at 20:25, Mag Gam <magawake@gmail.com> wrote:
Anyone had a chance to try this? :-)
$ for a in `seq 1 3`; do \time dd bs=4K count=1M if=/dev/zero of=/dev/null; done
Hm, I also wrote a little script to go through more values of bs. The script is attached. To change what values to use for bs, change the variables $min, $step, and $max at the beginning of the script, which are the parameters for "seq". To get a graph of throughput versus bs, use it like so: $ ./pipespeed.sh > data [...] $ gnuplot plot "data" w l The highest I get is 8.5 GB/s with bs=22528 on 2.6.37. How can you get a speed of 30 GB/s? It seems excessively fast. :) Greetings, Henry
Hi.... On Sat, Jan 22, 2011 at 19:00, Henry Gebhardt <hsggebhardt@googlemail.com> wrote:
How can you get a speed of 30 GB/s? It seems excessively fast. :)
I'd like to comment it first :) You read it incorrectly...it's 3.0, not 30 :D
Greetings,
Henry
-- regards, Mulyadi Santosa Freelance Linux trainer and consultant blog: the-hydra.blogspot.com training: mulyaditraining.blogspot.com
On Sat, Jan 22, 2011 at 11:30:52PM +0700, Mulyadi Santosa wrote:
On Sat, Jan 22, 2011 at 19:00, Henry Gebhardt wrote:
How can you get a speed of 30 GB/s? It seems excessively fast. :)
I'd like to comment it first :) You read it incorrectly...it's 3.0, not 30 :D
I was referring to the very first email in the original thread: On Wed, 19 Jan 2011 22:36:14 -0500, Mag Gam wrote:
I get around 30Gb/sec
However, when I do this with 2.6.37 I get close to 5GB/sec
Hm, I just noticed, its a 'b' on the first line, but capitalized, 'B', later. Just a typo? Greetings, Henry
On Sun, Jan 23, 2011 at 03:09, Henry Gebhardt <hsggebhardt@googlemail.com> wrote:
I was referring to the very first email in the original thread:
Oh that, Mag's...alright now I understand.
Hm, I just noticed, its a 'b' on the first line, but capitalized, 'B', later. Just a typo?
Maybe...but AFAIK these days sometimes people confusingly interchangeably use "b" and "B", "G" and "Gi" without really knowing what they really mean. Of course, I don't mean Mag must be falling into the same trap. But assuming it's really "b", it means it is around 3.75 GiB/s... that could be lower if he refer to "G", not "Gi"... -- regards, Mulyadi Santosa Freelance Linux trainer and consultant blog: the-hydra.blogspot.com training: mulyaditraining.blogspot.com
Hi again :) On Sat, Jan 22, 2011 at 19:00, Henry Gebhardt <hsggebhardt@googlemail.com> wrote:
The highest I get is 8.5 GB/s with bs=22528 on 2.6.37.
Thanks for sharing your script. Looking at the gnuplot output (again, thanks for the gnuplot quick hint :) ), I saw it is capped around 7.26 GiB/s, with block size around 300,000-400,000 (that means 293 - 391 KiB, right?) It's 2.6.36 btw... I think major factor that involve here, but not limited to them, are: - L1/L2/L3 cache size - slab allocator we use - current memory fragmentation level - read ahead size -- regards, Mulyadi Santosa Freelance Linux trainer and consultant blog: the-hydra.blogspot.com training: mulyaditraining.blogspot.com
participants (2)
-
Henry Gebhardt -
Mulyadi Santosa