[Beowulf] One network, or two?

Joe Landman landman at scalableinformatics.com
Tue Sep 23 19:01:22 EDT 2008



Prentice Bisbal wrote:
> Alan Ward wrote:
>> Good day.
>>
>> I have been reading the ongoing discussion on network usage with some
>> interest, mainly because in all (admittedly very small, 4 to 8 node)
>> clusters we have set up so far, we have always gone with doubling the
>> network. Nowadays we mostly run a 100 MBit/s "el cheapo" FastEthernet
>> for control, NFS and monitoring, while the faster Gigabit is exclusively
>> for MPI. Applications are CFD, with various levels of granularity.
>>
>> Anybody care to comment?
>>
>> -Alan
> 
> My new cluster, which is still in labor, will have InfiniBand for MPI,
> and we have 10 Gb ethernet switches for management/NFS, etc. The nodes
> only have 1 Gb ethernet, so it will be effectively a 1 Gb network.
> 
> I'm also curious as to whether the dual networks are overkill, and if
> using a slower network for I/O will cause the system to be slower than
> doing all traffic over IB, since I/O will be slower and cause the nodes
> to wait longer for these ops to finish.

Obviously YMMV, but in tuning systems, and seeing where bottlenecks are, 
we look for obvious things in the network design

1) poorly designed (office quality, usually uplinked/oversubscribed) 
networks ... have an interesting effect when you see two 128 port 
gigabit switches connected together with a 1 or 2 gigabit, or even 10 
GbE link.  You see this plateau in MPI scalability.  Yes, from a real 
customer case.  :(

2) I/O:  after burning incense to the daemons of low latency, the next 
big area we see is (curiously enough) I/O bandwidth.  I can't begin to 
elaborate on how many times I have seen a big expensive shiny new 
cluster with an absolutely terrible I/O design.  Usually starting with 1 
GbE connected NAS for 32 or more nodes.  Well, ok, YMMV, but when you 
run lots of Gaussian jobs which hammer on the NFS over this, you are 
going to experience pain.  And no, this is not where you stick the 
NetApps.  Yes, from several real customer cases.  :( :(

No, you don't need to perform ritual incantations to make the I/O go 
faster.  You just need good hardware and good design to get the data to 
the hardware.

1GbE may be great for EP jobs that occasionally write to disk.  But if 
your units are going to hammer on disk, you are going to need to check 
your IO design, and make sure it scales.  Yes, we are biased, we 
strongly believe in good I/O systems.


-- 
Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics LLC,
email: landman at scalableinformatics.com
web  : http://www.scalableinformatics.com
phone: +1 734 786 8423
fax  : +1 734 786 8452
cell : +1 734 612 4615

_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list