Charm++ parallel programming system

["Galvez Garcia, Juan Jose" <jjgalvez AT illinois.edu>]

Hi Chris,

 

Like you said, TopoManager provides detailed info for Cray and BG/Q topologies (like the shape of the topology, distance in network links between processors/nodes, and on what physical node each processor resides in).

Outside of these architectures, TopoManager can also be queried for the number of physical nodes and on which node each processor is in (this should work on most or all architectures).

 

We don’t currently have a way to query latency and bandwidth information in Charm++ at runtime. We do have a measurement-based approach which is used with Projections (Charm++’s performance analysis tool). With this approach, after running an application with tracing enabled, you can extract certain stats like message latency and bandwidth. This could be extended to provide this information at runtime. Ronak from our group (cc’ed) can give you more information about this functionality if you are interested.

 

-Juan

 

From: Chris Wailes
Sent: Tuesday, May 23, 2017 5:17 PM
To: charm
Subject: [charm] Querying Hardware Topology Information

 

I would like to build a topology graph for arbitrary machines, down to the processing element level, that includes latency and bandwidth information.  From the example programs and the existing load balancing strategies I found the TopoManager class, but it appears that this can only be used if the machine in question is a Cray or IBM BlueGene machine.  In addition, it doesn't appear to have the bandwidth or latency information I need.

Is there another API or tool I can use to obtain this information, or do I need to discover these properties outside of Charm++?

- Chris