[ofa-general] desired netperf mods?

[Rick Jones]

Scott Weitzenkamp (sweitzen) wrote:
> Rick,
> 
> SDP is easy to get with netperf by running "LD_PRELOAD=libsdp.so
> netperf/neterver", so in my opinion SDP is already covered.

The one and IMO very big worry I have about using LD_PRELOAD is that it 
does not change the netperf test banner.  So, it leaves one very 
vulnerable to cut-and-paste errors and not realizing that something was 
SDP when it really was.

The advantage then (again IMO) of an explicit SDP test is that the test 
banners will be right.

> 
> Here's a list of things I'd like to see from netperf, in priority order:
> 
> 1) IP multicast
> 2) Test that uses multiple concurrent sockets at the same time
> 3) RDS

The pt-pt nature of netperf (what I will often call netperf2) has often 
made me wonder if IP multicast in netperf2 made much sense - there would 
be one multicast sender and one multicast receiver.  If there is indeed 
value there, sure, IP/UDP multicast then.

Otherwise, I think that 1 and 2 are best served with netperf4.  Netperf4 
is where I am trying to address things not well suited to the pt-pt 
nature of netperf2 - multicast would be one of those, and multiple, 
concurrent sockets (with explicit synchronization of the measurment 
interval...) is another.

Netperf4 of course needs quite a bit of work/help.  FWIW it is also 
where I was able to get a "netperf" released under the GPL rather than 
the old "netperf license" from the HP Legal types ca 1993.

Netperf4 current mainline:  http://www.netperf.org/svn/netperf4/trunk

has a current, pretty much known to function set of bits.

Netperf4 branch: http://www.netperf.org/svn/branches/gobject_migration/

is where I'm trying to get a properly event driver netperf going to 
allow both interactive and batch benchmarking

I sometimes joke (well,perhaps half joke) that netperf4 is the 
eierlegendwolmilchsau (sp) version of netperf, or I call it "Shimmer" 
(being both a harness and a set of benchmark suites)

> I hadn't thought about netperf and uDAPL before, but certainly there
> would be value in adding uDAPL, MPI, and even filesystem benchmarks.  

As it happens, netperf4 does have a disc suite which was added by 
someone wanting to measure both networking and mass storage 
simultaneously on "combo" cards.

rick jones

> 
> Scott Weitzenkamp
> SQA and Release Manager
> Server Virtualization Business Unit
> Cisco Systems
>  
> 
> 
>>-----Original Message-----
>>From: general-bounces at lists.openfabrics.org 
>>[mailto:general-bounces at lists.openfabrics.org] On Behalf Of Rick Jones
>>Sent: Thursday, April 12, 2007 1:58 PM
>>To: general at lists.openfabrics.org
>>Subject: [ofa-general] desired netperf mods?
>>
>>Folks -
>>
>>I've only just started doing nefarious things like netperf over IPoIB 
>>and was wondering if "things" (for some definition of "things" in the 
>>context of Infiniband/whatnot) were stable-enough that some 
>>additional 
>>netperf tests would be warranted?  Some web searching 
>>suggests there may 
>>be AF_SDP (or perhaps AF_INET_SDP) explicit SDP support in places - 
>>would a nettest_sdp.c test suite be indicated?  How about uDAPL?
>>
>>Taps with a clue-bat most welcome - please be gentle :)
>>
>>happy benchmarking,
>>
>>rick jones
>>mr netperf
>>_______________________________________________
>>general mailing list
>>general at lists.openfabrics.org
>>http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
>>
>>To unsubscribe, please visit 
>>http://openib.org/mailman/listinfo/openib-general
>>