[distcc] Local vs. NFS lock files.

Donohue, Michael mdonohue at paypal.com
Wed Mar 30 19:49:15 GMT 2005


So I used our "business logic" archive to test performance of having
locks on a local disk verses on an NFS mount.

I used -j 15 to build the archive, and it consists of roughly 300 object
files.
I altered the DISTCC_DIR environment variable to point to either an NFS
mounted directory, or the local disk.    The tests were run late at
night, when no other developers were compiling.  Times are in minutes
and seconds.

Running with local locks, I get the first time of 5:48.  This is to load
up the cache consistently.
Then running with NFS locks, I get times of 5:36 and 5:39.   
Then, returning to local locks I get 6:49, and 5:43.

I'm not really seeing a benefit either way.  I should also note that IO
performance to the NFS mount is better than IO performance to the local
disk, due to gigabit Ethernet, and a hefty memory cache on the Network
Appliance.   The local disk can handle sustained writes of about 40
MB/second.   The NetApp can handle 100MB/sec sustained for at least 6GB.
I wouldn't expect IO performance to affect locking so much as latency
though, and I don't really know how the latency figures compare.

This is enough to convince me that NFS locking isn't hurting us at
PayPal, anyway.   What exactly are the issues that arise elsewhere?

Michael


More information about the distcc mailing list