On Mon, Aug 30, 2010 at 11:58 PM, Bart Van Assche <[email protected]> wrote: > On Tue, Aug 31, 2010 at 1:58 AM, Chris Worley <[email protected]> wrote: >> >> On Mon, Aug 30, 2010 at 5:46 PM, Chris Worley <[email protected]> wrote: >> > Hint: >> > >> > SCST rev 1901 ends compatibility w/ Windows; r1900 works fine. >> >> srpt_autodetect_cred_req was added in r1901, with default "Y". >> Setting it to "N"... or in later revs, "1", allows Windows to work >> again. > > (resending as plain text) > > That means there is a serious bug in the Windows SRP initiator. > Setting srpt_autodetect_cred_req=Y causes ib_srpt to test whether the > initiator supports SRP_CRED_REQ by sending a single SRP_CRED_REQ > information unit to the SRP initiator. Support for SRP_CRED_REQ is > required by the SRP (draft) standard and should not cause an initiator > to stop functioning properly.
Bart, I was a bit confused earlier: the default is "Y", it needs to be set to "N" (or, in current revs, it changed from logical to integer, so "0") to be compatible w/ any of the Windows SRP initiator drivers (back to 2.0.5). How does setting it to "0"/"N" effect performance? With this work-around, I seem to be peaking at 200K random 4KB IOPS in Windows... I was expecting a whole lot more (the target is capable of ~700K random 4KB IOPS in Linux testing, and recent papers suggest Windows is much faster than Linux as an initiator... at least in FC and iSCSI tests). These are initial results using two IB links between the target and initiator; I still need to play with the configuration a bit to assure I'm not doing anteing wrong. Might this setting alone account for the poor performance? Thanks, Chris > > Bart. > _______________________________________________ ofw mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ofw
