[linux-cifs-client] Some benchmarks and ideas

[Luis Claudio R. Goncalves]

Hi!

Playing with cifs 0.99a, samba3 and dbench I got the following results:

[root at gus sandbox]# dbench 1
1 clients started
   0     62477  3.68 MB/sec
Throughput 3.68239 MB/sec 1 procs
[root at gus sandbox]#
[root at gus sandbox]# dbench 2
2 clients started
   0     62477  4.14 MB/secc
Throughput 4.14458 MB/sec 2 procs
[root at gus sandbox]# dbench 4
4 clients started
   4       399  0.03 MB/sec

The system stall right at the beginning and after 10 minutes you see the
same data on screen and several error messages in the log. The most common
messages where "Send error on write" and "Send error on Close".

Repeating the test with dbench 5 our more iterations I always get the same
result but usually a bit faster:

[root at gus sandbox]#
[root at gus sandbox]# dbench 5
5 clients started
write failed on handle 13786
   5        47  0.00 MB/sec

I repeated these tests form dbench 1 to dbench 10 (the whole set of tests)
10 times and the results are always the same. More than 3 iterations, the
cifs mount stalls.

Besides the numbers meaning a higher load to the system (dbench 10 means
more than 10x the load of dbench1), in the point where the system stalls,
right at the begining, it sounds like the number of processes actively
writing to a cifs mount is the trigger to some bug.

I haven't dived deeply in the internals of cifsvfs, but it can be a good
thing to look for.

Regards,
Luis
-- 
[ Luis Claudio R. Goncalves                  lclaudio at conectiva.com.br ]
[ Fingerprint:   4FDD B8C4 3C59 34BD 8BE9  2696 7203 D980 A448 C8F8    ]
[ Msc has come!!!! - Conectiva HA Team - Gospel User - Linuxer - !Java ]
[ Fault Tolerance - Real-Time - Distributed Systems - IECLB - IS 40:31 ]
[ LateNite Programmer        --  My Utmost for His Highest  --         ]