Re: [Myricom help #8298] smpi_post_send_ok_to_send_out got corrupted handle

Hi Patrick,

    thanks. Be sure to apply this patch too. pkt_p->len should be zeroed.

-----------------------------------------------------------------------

--- mpich-1.2.1..7/mpid/ch_gm/smppriv.c.orig Tue Nov 27 14:36:33 2001
+++ mpich-1.2.1..7/mpid/ch_gm/smppriv.c Tue Nov 27 14:37:39 2001
@@ -594,7 +594,7 @@
pkt_p->mode = MPID_PKT_FLOW;
pkt_p->send_id = send_handle_id;
pkt_p->recv_id = recv_handle_id;
- pkt_len = 0;
+ pkt_p->len = 0;
shandle->gm.current_expected++;

DEBUG_PRINT_SMP_SEND_PKT("S Sending in the send_queue", pkt_p);

-------------------------------------------------------------------------

    Patrick, we've tried using mpich-1.2.1..7 & gm-1.5. Even with those two
patches Linpack never completes. It's works, but does not produce any result
for the problems of size above 30000 during large amount of time. We've also
tried mpich-1.2.1..7 & gm-1.4 with the same results... I don't know what
happens... Now we're using mpich-1.2..8 & gm-1.4 and everything works fine
with problems of size up too 230000.

    So, basicaly mpich-1.2.1..7 does not work properly on Alphas... We need
to do more investigation on the purpose, and then possibly we will find the
solution. Any ideas?

Regards,
    Oleg.

----- Original Message -----
From: "Patrick Geoffray" <[email protected]>
To: "Oleg I. Vdovikin" <[email protected]>
Cc: <[email protected]>; <[email protected]>
Sent: Thursday, November 29, 2001 5:44 PM
Subject: Re: [myrinet] Re: [Myricom help #8298]
smpi_post_send_ok_to_send_out got corrupted handle


> Hi Oleg,
>
> "Oleg I. Vdovikin" wrote:
> >
> > Hi,
> >
> >     there still no response from Myricom Technical Support...
>
> Thanksgiving is one of the rare national hollidays in the US, so
> a lot of people used the occasion to take some vacation last week.
> I am sorry for the delay: your email has been forwarded to me
> but my email queue was quite impressive when I came back yesterday.
>
> There is 2 people (Ruth and Susan) working full time for the Myricom
> support: the processing time is usually pretty good, except when
> everybody is away as it was the case last week.
>
> >     BTW, attached please find two patches against both mpich-1.2..8 &
> > mpich-1.2.1..7 which fixes Alpha (and IA64 as seems) alignment &
> > MPID_PKT_FLOW packet length errors. Without these patches most programs
> > crashed during smp operations.
>
> Your analysis is exactely right. Offsets in the shared memory receive
> queue should always use the full type of the packet, not the sum of
> the fields. I cleaned some part of the code in this way a few
> months ago, for alignement issue on Alpha also.
>
> I did not look carefully in the flow control code at that time. I will
> replace all references to sizes of packets' fields (like sizeof(void *))
> by the reference to the full type (much easier when adding fields into
> packet types). I will also changed the SMPI_PKT_BASIC definition to shrink
> the size of some fields (mode and lrank can fit in 16 bits).
SMPI_PKT_BASIC
> would then be 8 bytes aligned (it does not matter if all the offsets
> use full type sizes) and the latency a little bit lower.
>
> >     Please consider applying this patches to the source tree.
>
> I have corrected the CVS tree. It will be on the next release (need
> to finish the MPD integration and the gm_send/gm_directed_send
> failover before to release).
> Myricom support will put the patch on the Web and an entry in the
> FAQ.
>
> This bug has escaped the regression test because the test checking the
> flow control was run only on x86. I will add it to the test suite on
> alpha/IA64.
>
> Thanks for the report.
>
> Patrick
>
> ----------------------------------------------------------
> |   Patrick Geoffray, Ph.D.      [email protected]
> |   Myricom, Inc.                http://www.myri.com
> |   Cell:  865-389-8852          685 Emory Valley Rd (B)
> |   Phone: 865-425-0978          Oak Ridge, TN 37830
> ----------------------------------------------------------
>