Dear Colleagues;
I am doing a fluent parallel job on a 7 node 80 core heterogeneous system with different core number and size and different RAM also.
When I schedule works by cores, my data reading takes a long time, about a day. and I prefer to cancel data reading.
But on node or socket it is completed very fast, the problem is my CPU load of nodes which is very low and some cores seems to be idle.
To be honest I couldn't digest the aim of PBS module of fluent.
How can I force nodes to give their complete core and ram to me or else lower data reading in core mode to a reasonable time?
Thanks in advance
It seems you have bottleneck somewhere other than CPU. What kind of network you have?(Infiniband, Gigabit, ...) how big is your data? how fast is your network?(latency).
Dear Kaveh
many thanks.
you have mentioned a very interesting thing.to be honest I am fresh man in this field.
I try on one node and cpu usage is more than 90 percent.
I use Gigabit network with 1gbps cat6 cables.
My ping pong latency show a warning on head node that it is low (51 is shown),
although other nodes show a ping pong latency about 50, only head node shows a warning after test.
what does it mean?
How much should It be and how it could be improved?
Your latency on node should be around 0.02 ms and from one node to the other on GB Ethernet around 0.4 ms. I don't know units you mentioned but if it is in ms it is very high. For your calculations you need low latency network. (domain decomposition need updates on boundaries)
- Badges
- Science topic