[Swift-devel] RAM needs for running a swift workflow

[Mike Wilde]

Thanks, Yong.  So, it looks like a safe number is that it should be <32MB RAM 
for a small workflow (eg, 10s of nodes max).  perhaps much less. Looks like 
RAM goes up near-linear with #nodes in workflow.  But we dont know what the 
"floor" of fixed-overhead for Swift, Karajan and CoG is.

At 32M, 32 users would use 1GB of RAM for user JVMs, theoretically.

This sounds pretty reasonable and affordable, if a 2GB host could handle 20+ 
student users without paging.

Mike

Yong Zhao wrote, On 3/12/2007 2:44 PM:
> I have some numbers using the big_diamond workflow:
> The format is of following:
> 
> Mem
> iteration_in_workflow time_to_typecheck_the_graph number_of_nodes_in_wokflow
> 
> big.kml:
> 1GB
> 5698 - 24m54s - 165243
> 
> 512M
> 2832 - 13m4s - 82129
> 
> 256M
> 1393 - 6m29s - 40398
> 
> 128M
> 677 - 3m22s - 19634
> 
> 64m
> 317 - 1m38s - 9194
> 
> 32m
> 137 - 47s - 3974
> 
> Yong.
> 
> On Mon, 12 Mar 2007, Mike Wilde wrote:
> 
>> Does anyone have any data on how much memory is used for simple swift
>> workflows, and how that scales up?
>>
>> I ask because if we have numerous students in a workshop running swift on the
>> same submit host, we need to know how many it can handle before it consumes
>> all RAM and starts paging/swapping.
>>
>> A good thing to measure in the future for our application workflows (albeit
>> not the most critical for most applications).
>>
>> Alternatively - with the current command set, is there a practical way to have
>> many users share the same JVM to run small independent workflows?  (I assume
>> not, and that this might be a feature to consider in the distant future. Would
>> there be much savings for sharing a JVM, or is more of the memory overhead
>> per-workflow than per-JVM?)
>>
>> - Mike
>> _______________________________________________
>> Swift-devel mailing list
>> Swift-devel at ci.uchicago.edu
>> http://mail.ci.uchicago.edu/mailman/listinfo/swift-devel
>>
> 
> 

-- 
Mike Wilde
Computation Institute, University of Chicago
Math & Computer Science Division
Argonne National Laboratory
Argonne, IL   60439    USA
tel 630-252-7497 fax 630-252-1997