[ofa-general] RE: initial set of "direct" SDP tests in netperf

[Scott Weitzenkamp (sweitzen)]

Rick, I still think this is unecessary copy-paste code duplication, just
so the report says SDP instead of TCP.  Part of the advantage of
libsdp.so is you can use SDP w/o having to recode your application.

Do you really want to maintain duplicate code for all the
TCP_STREAM-vs-SDP_STREAM, etc. tests?  Yuck!

Scott Weitzenkamp
SQA and Release Manager
Server Virtualization Business Unit
Cisco Systems
 

> -----Original Message-----
> From: Rick Jones [mailto:rick.jones2 at hp.com] 
> Sent: Wednesday, April 25, 2007 2:23 PM
> To: general at lists.openfabrics.org
> Cc: Scott Weitzenkamp (sweitzen); Roland Dreier (rdreier)
> Subject: initial set of "direct" SDP tests in netperf
> 
> > I guess that until it is resolved I'll just kludge around 
> it with my own 
> > define.
> 
> Soo, I did a bunch of cut and paste in netperf, where I take what 
> getaddrinfo() returns and replace the ->ai_family with 
> AF_INET_SDP and 
> set the ->ai_protocol to 0 (a guess since there isn't much in 
> the way of 
> docs on the subject I could find).
> 
> I have implemented the SDP_STREAM, SDP_MAERTS and SDP_RR 
> tests thusfar, 
> they are enabled via a configure option of --enable-sdp .  
> You can grab 
> the bits from the top of trunk of the netperf2 repository at:
> 
> http://www.netperf.org/svn/netperf2/trunk/
> 
> I've done some initial, cursory testing with the bits that ship with 
> RHEL5 IA64 - I seem to have some sort of overlapping problem 
> still even 
> using the install.sh that purports to remove previous bits - and I've 
> tried the rpm -e command roland (?) posted a few days ago - 
> it says that 
> none of those things are installed.  I tried to modprobe ib_sdp (some 
> additional guesswork) and got symbol version mismatches. Still, even 
> with OFED 1.2rc2 removed there are ib_mumble modules loaded by RHEL5, 
> and .ko's in the standard (?) modules place.  So clearly I have some 
> remaining clueless noob issues wrt proper installation of 1.2 bits :(
> 
> happy benchmarking,
> 
> rick jones
> 
> BTW, speaking of cluelessness - you will probably have an 
> initial make 
> failure involving netperf_version.h - I've got something 
> still slightly 
> botched there I've not been able to figure-out just yet (make and 
> autotools aren't exactly my forte).  You can work around that 
> by cd'ing 
> to src/ and doing a "make netperf-version.h" and then go back 
> up and to 
> themake again.
> 
> PPS - the only remaining TCP tests of note I could bring-over 
> would be:
> 
> TCP_SENDFILE - can one use sendfile() against an SDP socket?
> TCP_CRR      - like TCP_RR but includes time to call connect()
> TCP_CC       - like TCP_CRR, but without the RR :)
> 
> feedback on which of those, if any, would be of interest 
> would be most 
> welcome.
> 
> After that I suppose RDS would be next on the list?  Is that 
> "real" at 
> this point?  Pointers on programming to it would be welcome.
>