[FLASH-USERS] flash network effects

[Tomasz Plewa]

Brock -

Many codes will scale well provide the interconnect is sufficiently  
fast or, conversely, that processing elements are relatively slow. The  
slow network asks for limiting the off-processor (off-node)  
communication and maximizing the amount of local calculations. Plus  
the excellent scaling for FLASH can be obtained only in a limited set  
of physics-devoid cases (see below).

No cure is readily available, but there are a few aspects worth considering:

(1) Is your problem well-balanced? The more physics you include, the  
more variation in processor usage will likely be observed.

(2) Is your algorithm simple enough? Yeah, hydro scales well. But  
hydro plus multigrid does not.

(3) Is the problem size large enough or communication plays  
significant role already for small pools of PEs?

(4) Have you tried any weak scaling tests for your particular problem?  
So does your problem scale "in theory"? (1 - 2 - 4 - 8 processors  
first...)

And there are likely a few more...

Greetings -

Tomek
-- 
School of Computational Science, DSL 443       ph:  850.644.1959
Florida State University                       fax: 850.644.0098
Tallahassee, FL 32306                          web: people.scs.fsu.edu/~tomek/
--



Brock Palen wrote:

> What really affects how many nodes a flash run can use?
> Also the notion that flash 'scales to hundreds of cpus' does that only
> apply to fast low latency networks.  Going from just 1 to 2 nodes (8 to
> 16 cpus)i see system time jump to 50% of cpu time.
>
> Do uses use gig-e at all? This is all we have avilable right now.  Any
> tricks for getting flash to run well on gig-e ?
>
>
> Brock Palen
> Center for Advanced Computing
> brockp at umich.edu
> (734)936-1985
>
>
>
> -- 
> This message has been scanned for viruses and
> dangerous content by MailScanner, and is
> believed to be clean.



----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.


-- 
This message has been scanned for viruses and
dangerous content by MailScanner, and is
believed to be clean.