[Beowulf] [External] anyone have modern interconnect metrics?

Prentice Bisbal pbisbal at pppl.gov
Sat Jan 20 02:39:58 UTC 2024


> Also ease of use with open-source products like OpenMPI,

I don't see this being an issue. OpenMPI will detect your different 
interconnects and start with the fastest and work its way to the 
slowest. OpenMPI has always "just worked" for me, regardless of the 
network. The only issue is if it doesn't find IB, it will issue a 
warning that it's using Ethernet and that's probably not what you want, 
but that's easy to turn off in the central config file.

> But vendors seem to think that high-end ethernet (100-400Gb) is 
> competitive... 
Call me a cynical old codger, but I would not be surprised if that's 
more profitable for them, or they have other incentives to promote 
Ethernet instead of IB. Or if you prefer Hanlon's razor, maybe they just 
don't know squat about IB so are selling you what they know.

>
> Yes, someone is sure to say "don't try characterizing all that stuff -
> it's your application's performance that matters!"  Alas, we're a generic
> "any kind of research computing" organization, so there are thousands 
> of apps
> across all possible domains. 

<rant>

I agree with you. I've always hated the "it depends on your application" 
stock response in HPC. I think it's BS. Very few of us work in an 
environment where we support only a handful of applications with very 
similar characteristics. I say use standardized benchmarks that test 
specific performance metrics (mem bandwidth or mem latency, etc.), 
first, and then use a few applications to confirm what you're seeing 
with those benchmarks.

</rant>

> Another interesting topic is that nodes are becoming many-core - any 
> thoughts? 

Core counts are getting too high to be of use in HPC. High core-count 
processors sound great until you realize that all those cores are now 
competing for same memory bandwidth and network bandwidth, neither of 
which increase with core-count.

Last April we were evaluating test systems from different vendors for a 
cluster purchase. One of our test users does a lot of CFD simulations 
that are very sensitive to mem bandwidth. While he was getting a 50% 
speed up in AMD compared to Intel (which makes sense since AMDs require 
12 DIMM slots to be filled instead of Intel's 8), he asked us consider 
servers with LESS cores. Even with the AMDs, he was saturating the 
memory bandwidth before scaling to all the cores, causing his 
performance to plateau. For him, buying cheaper processors with lower 
core-counts was better for him, since the savings would allow us to by 
additional nodes, which would be more beneficial to him.

>
> Alternatively, are there other places to ask? Reddit or something less 
> "greybeard"? 

I've been very disappointed with the "expertise" on the HPC-related 
subreddits. Last time I lurked there, it seemed very amateurish/DIY 
oriented. For example, someone wanted to buy all the individual 
components and build assemble their own nodes for an entire cluster at 
their job. Can you imagine? Most of the replies were encouraging them to 
do so....

  You might want to join the HPCSYSPROS Slack channel and ask there. HPC 
SYSPROS is an ACM SIG for HPC system admins that runs workshops every 
year at SC. click on the "Get Involved" link on this page:

https://sighpc-syspros.org/

--
Prentice


On 1/16/24 5:19 PM, Mark Hahn wrote:
> Hi all,
> Just wondering if any of you have numbers (or experience) with
> modern high-speed COTS ethernet.
>
> Latency mainly, but perhaps also message rate.  Also ease of use
> with open-source products like OpenMPI, maybe Lustre?
> Flexibility in configuring clusters in the >= 1k node range?
>
> We have a good idea of what to expect from Infiniband offerings,
> and are familiar with scalable network topologies.
> But vendors seem to think that high-end ethernet (100-400Gb) is 
> competitive...
>
> For instance, here's an excellent study of Cray/HP Slingshot (non-COTS):
> https://arxiv.org/pdf/2008.08886.pdf
> (half rtt around 2 us, but this paper has great stuff about 
> congestion, etc)
>
> Yes, someone is sure to say "don't try characterizing all that stuff -
> it's your application's performance that matters!"  Alas, we're a generic
> "any kind of research computing" organization, so there are thousands 
> of apps
> across all possible domains.
>
> Another interesting topic is that nodes are becoming many-core - any 
> thoughts?
>
> Alternatively, are there other places to ask? Reddit or something less 
> "greybeard"?
>
> thanks, mark hahn
> McMaster U / SharcNET / ComputeOntario / DRI Alliance Canada
>
> PS: the snarky name "NVidiband" just occurred to me; too soon?
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit 
> https://beowulf.org/cgi-bin/mailman/listinfo/beowulf


More information about the Beowulf mailing list