failures in smbtorture, dbench and friends

[acherry at pobox.com]

To throw even more fuel onto the fire, I've tried the following in a
few situations (smbtorture from Samba 2.0.7 release, all Samba servers
are using Samba 2.0.7)

smbtorture //testing/netbench -U user%pass -N 16 NBWNT

Client			Server			Results
------			------			-------
Sun Enterprise 4500	Sun Ultra 60		Many "write failed" errors,
4x400 MHz UltraSP2	360MHz / 512M		segfaults.
2GB physical mem

Sun Enterprise 4500	Self (loopback)		Same as above.

Sun Enterprise 4500	Compaq Proliant 6000	No "write failed" errors,
			NT Terminal Server SP5	test completes.
			Dual PII Xeon 400, 1GB

Indigo2/IMPACT R10K	Self (loopback)		Same sort of problems
256M						as the mentioned above.

All of the above setups work fine for 8 clients instead of 16.  If I
bump the number of clients up to 32, the WTS server test completes,
but generates a single "tdis failed (ERRDOS - ERRbadfid (Invalid file
handle.))"  error message. Also, for ALL of the WTS tests, I get a
bunch of "ctcxspwts2 rejected the session" errors at the beginning,
corresponding with "The server was unable to allocate a work item"
errors in the Event Log, although all of the clients do appear in
Server Manager when I check, so it's possible that some of the clients
fail to connect on the first try but succeed after a number of
attempts.

So it looks as if both the number of clients and the server on the
other end affect the success/failure of the tests.

Don't know if this info is useful, but I thought I'd provide it
anyway.

-Andrew

J. Robert von Behren writes:
 > 
 > To throw a bit more fuel (but unfortunately no further clues) on the
 > fire, I've seen the exact same behavior on a couple of Linux test
 > servers I've been playing with - lots of "nb_close: handle XXX was not
 > open" and "write failed on handle XXXX" errors, and correspondingly
 > bogus results.  I've seen this on several different Linux kernel
 > versions (in the 2.3.99 and 2.4.0-test series), and with both the stock
 > Samba 2.0.7 and a couple of snapshots from the main branch of the CVS
 > tree.
 > 
 > The problem appears somewhat random, although it seems to happen more
 > often with larger numbers of clients in smbtorture.  I'd be happy to do
 > some additional exploration if anyone has ideas of things to try.  (I'll
 > have only intermittent email contact until 8/27, however, so I may not
 > be able to do much until after then.)
 > 
 > Best regards,
 >