[Savannah-hackers-public] vc load

[Karl Berry]

A few minutes ago, I noticed that a simple cvs update in a small
directory for a project on savannah was not completing.

Fortunately, I already had a shell on vcs.  It showed that loggerhead
(bzr web browsing) was consuming huge resources:

top - 15:24:32 up 251 days,  8:01,  1 user,  load average: 38.35, 20.86, 15.91
  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND           
 1488 www-data  39  19 2429m 1.3g 2724 S   82 21.8   7847:09 python /usr/src/lo

(The full program for pid 1488 was
python /usr/src/loggerhead-1.18.1/serve-branches --host=127.0.0.1 /srv/bzr 
--prefix=/lh --allow-writes)

I killed it and restarted it.

Apparently loggerhead is started once by rc and lives until the system
reboots (251 days and counting).  I suppose some kind of balance finally
got tipped.  Sure looking forward to Bob's reboot program completing.

There may have been contention with git, since the next three processes
were:
  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND           
12573 nobody    30  10 74336  35m  33m D    7  0.6   0:00.36 git pack-objects -
11886 nobody    30  10  413m  84m  28m S    3  1.4   0:12.84 git pack-objects -
12443 nobody    30  10 75780  39m  21m D    2  0.6   0:02.08 git pack-objects -

Don't know what to make of that.  The git commands all went back to
inetd but it wasn't obvious which project they were associated with, if
any.  They completed reasonably quickly (a few minutes) after loggerhead
was killed, but still.


Finally, in /etc/xinetd.d/*, I noticed that all the rlimit_cpu values
were commented out, as in:
service git  # or bzr or cvs or whatever
{
        disable         = no
        #rlimit_cpu      = 20

I've never used rlimit_cpu, but it sure sounds like a good idea.  Did it
prove problematic in practice?

k