I needed to recompress some files from .bz2 to .gz so I setup a script to
do
bunzip2 -c $file.bz2 |gzip -9 >$file.gz
I expected that the two CPU heavy processes would end up on different
cpu's and spend a little time shuffling data between the two cpu's on a
system (dual core opteron)
however, instead what I find is that each process is getting 50% of one
cpu while the other cpu is 97% idle.
David Lang
> I needed to recompress some files from .bz2 to .gz so I setup a script to
> do
>
> bunzip2 -c $file.bz2 |gzip -9 >$file.gz
>
> I expected that the two CPU heavy processes would end up on different
> cpu's and spend a little time shuffling data between the two cpu's on a
> system (dual core opteron)
>
> however, instead what I find is that each process is getting 50% of one
> cpu while the other cpu is 97% idle.
That would only be possible if the compression/decompression block size is
small compared to the maximum pipe buffer size. I suspect the reverse is the
case.
It would be interesting to write an intermediate process that basically
enlarged the pipe buffers and see if that changed anything. Basically, the
intermediate process would allocate a large buffer (16MB or so) and fill it
from 'bunzip2' while draining it to 'gzip' in a non-blocking way (unless the
buffer was full/empty, of course).
DS
On Thu, 3 May 2007, David Schwartz wrote:
>> I needed to recompress some files from .bz2 to .gz so I setup a script to
>> do
>>
>> bunzip2 -c $file.bz2 |gzip -9 >$file.gz
>>
>> I expected that the two CPU heavy processes would end up on different
>> cpu's and spend a little time shuffling data between the two cpu's on a
>> system (dual core opteron)
>>
>> however, instead what I find is that each process is getting 50% of one
>> cpu while the other cpu is 97% idle.
>
> That would only be possible if the compression/decompression block size is
> small compared to the maximum pipe buffer size. I suspect the reverse is the
> case.
>
> It would be interesting to write an intermediate process that basically
> enlarged the pipe buffers and see if that changed anything. Basically, the
> intermediate process would allocate a large buffer (16MB or so) and fill it
> from 'bunzip2' while draining it to 'gzip' in a non-blocking way (unless the
> buffer was full/empty, of course).
hmm, how about
bunzip2 -c $file.bz2 |dd bs=8m |gzip -9 >$file.gz
should that work?
David Lang
On Thu, 3 May 2007, David Schwartz wrote:
>> I needed to recompress some files from .bz2 to .gz so I setup a script to
>> do
>>
>> bunzip2 -c $file.bz2 |gzip -9 >$file.gz
>>
>> I expected that the two CPU heavy processes would end up on different
>> cpu's and spend a little time shuffling data between the two cpu's on a
>> system (dual core opteron)
>>
>> however, instead what I find is that each process is getting 50% of one
>> cpu while the other cpu is 97% idle.
>
> That would only be possible if the compression/decompression block size is
> small compared to the maximum pipe buffer size. I suspect the reverse is the
> case.
I'm still running into this problem in various forms
is there an easy way to change the maximum pipe buffer size? (including a
simple change to the kernel source, I do compile my own kernels)
> It would be interesting to write an intermediate process that basically
> enlarged the pipe buffers and see if that changed anything. Basically, the
> intermediate process would allocate a large buffer (16MB or so) and fill it
> from 'bunzip2' while draining it to 'gzip' in a non-blocking way (unless the
> buffer was full/empty, of course).
> On Thu, 3 May 2007, David Schwartz wrote:
>
> >> I needed to recompress some files from .bz2 to .gz so I setup
> a script to
> >> do
> >>
> >> bunzip2 -c $file.bz2 |gzip -9 >$file.gz
> >>
> >> I expected that the two CPU heavy processes would end up on different
> >> cpu's and spend a little time shuffling data between the two cpu's on a
> >> system (dual core opteron)
> >>
> >> however, instead what I find is that each process is getting 50% of one
> >> cpu while the other cpu is 97% idle.
> >
> > That would only be possible if the compression/decompression
> block size is
> > small compared to the maximum pipe buffer size. I suspect the
> reverse is the
> > case.
>
> I'm still running into this problem in various forms
>
> is there an easy way to change the maximum pipe buffer size? (including a
> simple change to the kernel source, I do compile my own kernels)
No. Changing the size will not do what you want it to do since that only
tells the kernel what the size is, it does not determine what it is.
> > It would be interesting to write an intermediate process that basically
> > enlarged the pipe buffers and see if that changed anything.
> > Basically, the
> > intermediate process would allocate a large buffer (16MB or so)
> > and fill it
> > from 'bunzip2' while draining it to 'gzip' in a non-blocking
> > way (unless the
> > buffer was full/empty, of course).
It is not particularly hard to write such a process. I have a proxy that I
can easily tweak to do this. I'm going to give it a shot and see if it
helps.
DS
> > >> bunzip2 -c $file.bz2 |gzip -9 >$file.gz
So here are some actual results from a dual P3-1Ghz machine (2.6.21.1,
CFSv9). First lets time each operation individually:
$ time bunzip2 -k linux-2.6.21.tar.bz2
real 1m5.626s
user 1m2.240s
sys 0m3.144s
$ time gzip -9 linux-2.6.21.tar
real 1m17.652s
user 1m15.609s
sys 0m1.912s
The compress was the most complex (no surprise there) but they are close
enough that efficient overlap will definitely affect the total wall time. If
we can both decompress and compress in 1:17, we are optimal. First, let's
try the normal way:
$ time (bunzip2 -c linux-2.6.21.tar.bz2 | gzip -9 > test1)
real 1m45.051s
user 2m16.945s
sys 0m2.752s
1:45, or 1/3 over optimal. Now, with a 32MB non-blocking cache between the
two processes ('accel' creates a 32MB cache and uses 'select' to fill from
stdin and empty to stdout without blocking either direction):
$ time (bunzip2 -c linux-2.6.21.tar.bz2 | ./accel | gzip -9 > test2)
real 1m18.361s
user 2m19.589s
sys 0m6.356s
Within testing accuracy of optimal.
So it's not the scheduler. It's the fact that bunzip2/gzip have inadequate
input/output buffering. I don't think it's unreasonable to consider this a
defect in those programs.
DS
David Schwartz wrote:
>>>>> bunzip2 -c $file.bz2 |gzip -9 >$file.gz
>
> So here are some actual results from a dual P3-1Ghz machine (2.6.21.1,
> CFSv9). First lets time each operation individually:
>
> $ time bunzip2 -k linux-2.6.21.tar.bz2
>
> real 1m5.626s
> user 1m2.240s
> sys 0m3.144s
>
>
> $ time gzip -9 linux-2.6.21.tar
>
> real 1m17.652s
> user 1m15.609s
> sys 0m1.912s
>
> The compress was the most complex (no surprise there) but they are close
> enough that efficient overlap will definitely affect the total wall time. If
> we can both decompress and compress in 1:17, we are optimal. First, let's
> try the normal way:
>
> $ time (bunzip2 -c linux-2.6.21.tar.bz2 | gzip -9 > test1)
>
> real 1m45.051s
> user 2m16.945s
> sys 0m2.752s
>
> 1:45, or 1/3 over optimal. Now, with a 32MB non-blocking cache between the
> two processes ('accel' creates a 32MB cache and uses 'select' to fill from
> stdin and empty to stdout without blocking either direction):
>
> $ time (bunzip2 -c linux-2.6.21.tar.bz2 | ./accel | gzip -9 > test2)
>
> real 1m18.361s
> user 2m19.589s
> sys 0m6.356s
>
> Within testing accuracy of optimal.
>
> So it's not the scheduler. It's the fact that bunzip2/gzip have inadequate
> input/output buffering. I don't think it's unreasonable to consider this a
> defect in those programs.
>
They are hardly designed to optimize this operation...
For a tunable buffer program allowing the buffer size and buffers in the
pool to be set, see http://www.tmr.com/~public/source program ptbuf. I wrote it
as a proof of concept for a pthreads presentation I was giving, and it
happened to be useful.
--
Bill Davidsen <[email protected]>
"We have more to fear from the bungling of the incompetent than from
the machinations of the wicked." - from Slashdot