03-07-2014 01:18 PM
Iam trying to run the SmartXplorer tool using several custom strategies. I have a 8-node Linux cluster with 8-cores on each node. I am trying to run 64 jobs simultaneously. The Xilinx installation and the working directory are NFS mounted and accessible on all nodes. I am able to see all the jobs running in parallel in the begining. But as time progresses, the scheudling is very slow. Some nodes are just idle (top command shows no processes running) and all of a sudden, there are 8 build running. It doesnt march through all the strategies as quickly as expected and as quickly as it does in the begining. Any thoughts or suggestions would be helpful.
03-08-2014 07:12 AM
My experience with multicore processors is that you need to have enough memory (physical system RAM) to run all copies for that machine. i.e. if your design typically uses 2 GB during Map, then to run 8 processors you'd need at least 16 GB of system memory. Any less and you will not gain speed because you end up using the disk as virtual memory, and the memory access during map is pretty much random. This creates disk thrashing and the whole process slows down. For some versions of ISE, Map will exit with an error if there is not sufficient physical memory. So you might want to try to run fewer jobs per node to see if that helps.