[ofa-general] Is response time of GMP is more than SMP

[Hal Rosenstock]

Hi Sumit,

On Fri, 2008-06-27 at 19:28 +0530, Sumit Gaur - Sun Microsystem wrote:
> hi Hal,
> 
> Hal Rosenstock wrote:
> > 
> >>>Can you elaborate on the multiple sends ? Are they outstanding
> >>>concurrently ? Are they to the same destination or different ones ? Are
> >>>they from a single or multiple threads ?
> >>
> >>No they are sending sequentially(mutex enabled) no concurrency but timeout for 
> >>umad_recv is 100ms.
> > 
> > 
> > Can you try increasing that to see if there is some threshold where it
> > works more reliably ? Does it work better at say 200 msec (as you said
> > your rate was 4/sec) ? The default timeout used in the diags is 1 sec.
> 
> yes, I tried increasing it upto 3000ms but condition not improved much (it 
> *reduced* timeout failures but *no reduction* in recv_fail and this piling up 
> number of request per second too at client side).

Right; I'm not sure you can send the subsequent request until the former
one either completes or times out.

>  Also I tried same code on two 
> separate subnet and similar problem come across.

Is the second subnet any better/cleaner than the first ?

>  Could be a network issue but is 
> it a fact that GMP request takes much more time then SMP request (for ideal 
> subnet) ?

Other than SMPs being VL15 with no flow control and GMPs being data VL
(usually VL0) with flow control, there should be no difference for SMA
v. PMA requests AFAIK.

-- Hal

> > Looks like there are some issues here to debug in your subnet. It might
> > help to clear the counters and see what is actively going on to isolate
> > these issues. This could factor into those other errors you are seeing.
> > 
> > -- Hal
> > 
> >