DPDK usage discussions
 help / color / mirror / Atom feed
* [dpdk-users] Which coremask and "-n" (number of mem channels) should be used for a system with only "node1" (no "node0") and 2 sockets ?
@ 2016-07-15  6:51 Kevin Wilson
  2016-07-18 19:13 ` Matthew Hall
  0 siblings, 1 reply; 5+ messages in thread
From: Kevin Wilson @ 2016-07-15  6:51 UTC (permalink / raw)
  To: users

Hello,
I have a machine on which there is a bit non common setup: I have only
"node1" (no node0) under /sys/devices/system/node:
ls /sys/devices/system/node/
...
node1
...

I have 2 sockets on this machines, and 16 cores:

./tools/cpu_layout.py
============================================================
Core and Socket Information (as reported by '/proc/cpuinfo')
============================================================

cores =  [0, 1, 2, 3]
sockets =  [0, 1]

       Socket 0        Socket 1
       --------        --------
Core 0 [0, 8]          [4, 12]
Core 1 [1, 9]          [5, 13]
Core 2 [2, 10]         [6, 14]
Core 3 [3, 11]         [7, 15]

What should be the "-c" and "-n"  EAL paramaters that I should use ? I
 made several trials but
none of them succeeded, applications cannot start.

Please let me know if anything else is needed.

Regards,
Kevin

^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: [dpdk-users] Which coremask and "-n" (number of mem channels) should be used for a system with only "node1" (no "node0") and 2 sockets ?
  2016-07-15  6:51 [dpdk-users] Which coremask and "-n" (number of mem channels) should be used for a system with only "node1" (no "node0") and 2 sockets ? Kevin Wilson
@ 2016-07-18 19:13 ` Matthew Hall
  2016-08-16 17:41   ` Kevin Wilson
  0 siblings, 1 reply; 5+ messages in thread
From: Matthew Hall @ 2016-07-18 19:13 UTC (permalink / raw)
  To: Kevin Wilson; +Cc: users

On Fri, Jul 15, 2016 at 09:51:42AM +0300, Kevin Wilson wrote:
> Hello,
> I have a machine on which there is a bit non common setup: I have only
> "node1" (no node0) under /sys/devices/system/node:
> ls /sys/devices/system/node/
> ...
> node1
> ...

If I had to guess then either the BIOS is busted or the CPU is in the wrong 
socket or the RAM is in the wrong socket.

Either way all bets are off how this is going to pan out. Probably badly when 
it comes to getting the top performance.

Some exploration of the dmidecode output on the machine could be valuable.

Matthew.

^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: [dpdk-users] Which coremask and "-n" (number of mem channels) should be used for a system with only "node1" (no "node0") and 2 sockets ?
  2016-07-18 19:13 ` Matthew Hall
@ 2016-08-16 17:41   ` Kevin Wilson
  2016-08-16 18:05     ` Stephen Hemminger
  0 siblings, 1 reply; 5+ messages in thread
From: Kevin Wilson @ 2016-08-16 17:41 UTC (permalink / raw)
  To: Matthew Hall; +Cc: users

Thanks Metthew.
Indeed,  something seems broken in that machine,

Running
dmidecode
gives:
.....
Invalid entry length (0). DMI table is broken! Stop

Kevin


On Mon, Jul 18, 2016 at 10:13 PM, Matthew Hall <mhall@mhcomputing.net> wrote:
> On Fri, Jul 15, 2016 at 09:51:42AM +0300, Kevin Wilson wrote:
>> Hello,
>> I have a machine on which there is a bit non common setup: I have only
>> "node1" (no node0) under /sys/devices/system/node:
>> ls /sys/devices/system/node/
>> ...
>> node1
>> ...
>
> If I had to guess then either the BIOS is busted or the CPU is in the wrong
> socket or the RAM is in the wrong socket.
>
> Either way all bets are off how this is going to pan out. Probably badly when
> it comes to getting the top performance.
>
> Some exploration of the dmidecode output on the machine could be valuable.
>
> Matthew.

^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: [dpdk-users] Which coremask and "-n" (number of mem channels) should be used for a system with only "node1" (no "node0") and 2 sockets ?
  2016-08-16 17:41   ` Kevin Wilson
@ 2016-08-16 18:05     ` Stephen Hemminger
  2016-08-16 18:09       ` Matthew Hall
  0 siblings, 1 reply; 5+ messages in thread
From: Stephen Hemminger @ 2016-08-16 18:05 UTC (permalink / raw)
  To: Kevin Wilson; +Cc: Matthew Hall, users

On Tue, 16 Aug 2016 20:41:42 +0300
Kevin Wilson <wkevils@gmail.com> wrote:

> Thanks Metthew.
> Indeed,  something seems broken in that machine,
> 
> Running
> dmidecode
> gives:
> .....
> Invalid entry length (0). DMI table is broken! Stop
> 
> Kevin
> 
> 
> On Mon, Jul 18, 2016 at 10:13 PM, Matthew Hall <mhall@mhcomputing.net> wrote:
> > On Fri, Jul 15, 2016 at 09:51:42AM +0300, Kevin Wilson wrote:  
> >> Hello,
> >> I have a machine on which there is a bit non common setup: I have only
> >> "node1" (no node0) under /sys/devices/system/node:
> >> ls /sys/devices/system/node/
> >> ...
> >> node1
> >> ...  
> >
> > If I had to guess then either the BIOS is busted or the CPU is in the wrong
> > socket or the RAM is in the wrong socket.
> >
> > Either way all bets are off how this is going to pan out. Probably badly when
> > it comes to getting the top performance.
> >
> > Some exploration of the dmidecode output on the machine could be valuable.
> >
> > Matthew.  

What if you have a two socket machine, and node 0 is not populated?

Also does the kernel have NUMA support enabled?

Long ago there was a bug in DPDK where it got NUMA socket info from /proc/cpuinfo.
On some machines, the 'physical id' starts at 1 because value comes from BIOS.

^ permalink raw reply	[flat|nested] 5+ messages in thread

* Re: [dpdk-users] Which coremask and "-n" (number of mem channels) should be used for a system with only "node1" (no "node0") and 2 sockets ?
  2016-08-16 18:05     ` Stephen Hemminger
@ 2016-08-16 18:09       ` Matthew Hall
  0 siblings, 0 replies; 5+ messages in thread
From: Matthew Hall @ 2016-08-16 18:09 UTC (permalink / raw)
  To: Stephen Hemminger; +Cc: Kevin Wilson, users

On Tue, Aug 16, 2016 at 11:05:26AM -0700, Stephen Hemminger wrote:
> What if you have a two socket machine, and node 0 is not populated?

I did include that possibility. But most NUMA systems I have set up myself 
will not boot if Socket 1 is full and Socket 0 is empty. The early boot code 
generally assumes Socket 0 is valid when first initting itself.

> Also does the kernel have NUMA support enabled?

Good question, the vast majority of standard SMP kernels do have NUMA at this 
point I believe. It would maybe be good to see if the issue went away if the 
machine was intentionally run with a UP kernel.

> Long ago there was a bug in DPDK where it got NUMA socket info from 
> /proc/cpuinfo. On some machines, the 'physical id' starts at 1 because value 
> comes from BIOS.

Ouch... that's very evil of the BIOS. But also sadly not surprising.

Matthew.

^ permalink raw reply	[flat|nested] 5+ messages in thread

end of thread, other threads:[~2016-08-16 18:09 UTC | newest]

Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2016-07-15  6:51 [dpdk-users] Which coremask and "-n" (number of mem channels) should be used for a system with only "node1" (no "node0") and 2 sockets ? Kevin Wilson
2016-07-18 19:13 ` Matthew Hall
2016-08-16 17:41   ` Kevin Wilson
2016-08-16 18:05     ` Stephen Hemminger
2016-08-16 18:09       ` Matthew Hall

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).