On Nov 3, 2007 4:43 AM, Philip Martin <philip@codematters.co.uk> wrote:
> It depends on how fast the hardware generates entropy in comparison to
> how fast the testsuite runs. I used to use an older dual processor
> machine with multiple SCSI disks, multiple network cards, a separate
> video card, etc. and it ran the testsuite without stalling. I now use
> a newer machine with onboard video/LAN/SATA and a single disk and it
> runs the testsuite much faster but will stall when it runs out of
> entropy. These stalls are most visible when using ra_svn. The times
> you give for the tests are relatively slow compared to the times I get
> out of my new machine where basic_tests over ra_local or ra_dav runs
> in under 20 seconds.
This is a relatively recent Mac Mini (2x 1.8GHz; Core Duo T2400)
running Ubuntu 7.10. My MBP on Mac OS X is a 2.33 C2D. Neither are
old machines.
Are your 20 second tests run off a RAM drive or something? I'm
running off the local HDD and not doing parallel runs - straight 'make
check' out-of-the-box.
Here is 'time' output from a complete 'make check' run (46 minutes):
186.27user 135.30system 46:05.98elapsed 11%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (0major+13369289minor)pagefaults 0swaps
So, our tests are obviously not CPU-bound. I have a hunch you're
doing something with a RAM drive. Also, the tests completed in almost
identical time with /dev/urandom or /dev/random.
> Perhaps your machine is simply too slow to drain
> its entropy pool? On my machine ra_svn is as fast as the other layers
> when using /dev/urandom, under 20 seconds, but takes over 90 seconds
> using /dev/random.
>
> On Linux you can use
>
> cat /proc/sys/kernel/random/entropy_avail
>
> to see how much entropy is available, and ra_svn does drain it more
> noticeably than the other layers.
Again, there is no statistical difference between /dev/random and
/dev/urandom for me on ra_serf / ra_local. So, I doubt it's entropy
as urandom shouldn't be blocked by entropy.
If it's ra_svn, then if you are using the MD5-CRAM authentication,
then that would be a likely culprit. Using SASL with PLAIN would
avoid that - perhaps we should configure out testsuite to do that.
However, if it is just ra_svn, that wouldn't explain Karl seeing a
multi-factor speed-up on ra_dav or ra_local by changing the entropy
source. -- justin
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@subversion.tigris.org
For additional commands, e-mail: dev-help@subversion.tigris.org
Received on Sat Nov 3 16:49:29 2007