[Beowulf] NUMA zone weirdness
hearnsj at googlemail.com
Sun Dec 18 01:42:21 PST 2016
Just another quick thought on this.
The routine pio.c steps through all NUMA zones and checks they are online.
As we all know (yes, here come the war stories again... zzz) I used to
manage an Untraviolet which had 128 NUMA zones.
I guess I never saw it boot up with any NUMA zones disabled - and probably
I woud have had kittens if that happened.
But looks like the Omnipath driver would check all 128 are online before
ps. to make it a bit clearer - thats what clued me into this problem. I
tried to run a simple HPL on one of those nodes,
with Intel MPI and got back a Hydra failure. Checking through dmesg I saw
a failure in the loading of the hfi1 driver,
and had to look at the driver code which was causing this error message.
then homed on on pio.c looking at the online status.
Clearly in my case the NUMA zones were 0 and 2 - so looping 0,1 found
that 1 was offline!
On 16 December 2016 at 22:35, Peter St. John <peter.st.john at gmail.com>
> I notice an odd thing maybe someone more hardware cluefull could explain?
> Node 1 has 32GB (that is, 32 * 1024 MB) but Node 0 is an odd number (very
> odd, to me), 32673 is 95 MB short. It doesn't make sense to me that a bank
> of bad memory would be such a funny number short.
> On Fri, Dec 16, 2016 at 4:59 PM, Greg Lindahl <lindahl at pbm.com> wrote:
>> Wow, that's pretty obscure!
>> I'd recommend reporting it to Intel so that they can add it to the
>> descendants of ipath_checkout / ipath_debug. It's exactly the kind of
>> hidden gotcha that leads to unhappy systems!
>> -- greg
>> On Fri, Dec 16, 2016 at 03:52:34PM +0000, John Hearns wrote:
>> > Problem solved.
>> > I have changed the QPI Snoop Mode on these servers from
>> > ClusterOnDIe Enabled to Disabled and they display what I take to be
>> > behaviour - ie
>> > [root at comp006 ~]# numactl --hardware
>> > available: 2 nodes (0-1)
>> > node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11
>> > node 0 size: 32673 MB
>> > node 0 free: 31541 MB
>> > node 1 cpus: 12 13 14 15 16 17 18 19 20 21 22 23
>> > node 1 size: 32768 MB
>> > node 1 free: 31860 MB
>> > node distances:
>> > node 0 1
>> > 0: 10 21
>> > 1: 21 10
>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
>> To change your subscription (digest mode or unsubscribe) visit
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Beowulf