>Everything should get cleaned up when the app exits. Does the memory >get reclaimed after the test, or just increase during execution? Have >you looked at other tests, to see if they exhibit the same problem? >(I'll take a look at this today.)
A couple more data points with ndconn. I do see the memory increase for winverbs, so I'll try to track that down. Neither the ibal or winverbs providers are CPU limited. The client reports a greater conn rate than server side. For ibal, I saw a difference of 100 conn/sec, which equates to over 6000 more connections on the client side than the server side for the duration of the test. Some difference is understandable, but this seems high. Winverbs rates were 500 to 900 from run to run. Ibal rates were 1300 to 1400. I hacked winverbs to avoid the socket calls except for the ND listen routine. I didn't see any measurable difference with ndconn. Actually, with the calls left in, I usually had better rates (800 versus 500 for most non-socket runs), which doesn't make sense. Ibal to winverbs does not work in either direction. I thought I had tested this before, but apparently not. The problem is probably the difference in the port space that is used, which results in different service IDs. - Sean _______________________________________________ ofw mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ofw
