[SCore-users] Help: Too slow running any program

[Huiqun Zhou]

Hi, Bogdan,

Sorry for my delay of responding you. It seems that programs running in my Score environment
not only took long, long time between "Score-D connected" and the first line of output
displayed, but also spent more time on communication during computation. I tried a fluid
dynamics code on my cluster. For its serial version, it took about 2 seconds per each time step.
While for its parallel version, besides it took about 3 minutes at the beginning of the run, it took
4 to 6 seconds per time step even on a 7 compute node cluster!

The code I tried run very well on my previous cluster, which comprised of 4 compute nodes and 
a dedicated server. All machines of my previous cluster are P III 800MHz, 256 MB with Intel NIC.
My current cluster is comprised of 8 machines with P4 2.4GHz, 512MB and 3Com NIC, plus
D-Link DGE-550T Gigabit NIC (not in use yet).

I'm using D-Link's 8 port Gigabit Ethernet switch DGS-1008T.

Looking forward to your help.

----------------------------
Huiqun Zhou
Department of Earth Sciences
Nanjing University
China

e-mail: hqzhou at nju.edu.cn
Tel: 86(25)359-4664
FAX: 86(25)368-6016
Mobil: 13182856800
----------------------------

----- Original Message ----- 
From: "Bogdan Costescu" <bogdan.costescu at iwr.uni-heidelberg.de>
To: "Huiqun Zhou" <hqzhou at nju.edu.cn>
Cc: <SCore-users at pccluster.org>
Sent: Monday, July 21, 2003 10:04 PM
Subject: Re: [SCore-users] Help: Too slow running any program


> On Fri, 18 Jul 2003, Huiqun Zhou wrote:
> 
> > It even took 2 minutes or more to run the simple examples, hello and
> > cpi. Although the results are correct, it seems that it keeps looking
> > for something at the begining of the run.
> 
> I the beginning of the run the only part affected ? I mean, does it take a 
> lot of time before printing the first line of output, but afterwards 
> everthing works with normal speed ? What is the load of the nodes when 
> this happens ?
> 
> > only 3COM card is in use.
> 
> With Ethernet hardware, especially with FastEthernet one, care must be
> taken that the speed and duplex setting of the switch ports and cards are
> in harmony. The best solution is to set both the card (default with the
> 3c59x driver) and the switch port to autonegotiate/NWAY. Failure to do so 
> will provide a link with very high probability of packet loss which 
> drastically reduces the maximum data transfer rate that can be obtained as 
> packets have to be retransmitted. Obviously, any kind of parallel 
> computation that uses such a broken link would be very slow.
> 
> -- 
> Bogdan Costescu
> 
> IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
> Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
> Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
> E-mail: Bogdan.Costescu at IWR.Uni-Heidelberg.De
> 
> _______________________________________________
> SCore-users mailing list
> SCore-users at pccluster.org
> http://www.pccluster.org/mailman/listinfo/score-users
> 
>