Message boards : Number crunching : Test task
Message board moderation
Author | Message |
---|---|
Send message Joined: 16 Aug 15 Posts: 966 Credit: 1,211,816 RAC: 0 |
I suggest to implement a tick box in the preferences to enable a test task. This would consist of a carefully chosen job, and only one job(same on multi-core tasks). (as a precaution, this could automatically switch to "normal" after a period e.g 24h) As all normal jobs have different length, they cannot be compared. Purpose would be, to test different configurations for best speed. Comments? |
Send message Joined: 12 Sep 14 Posts: 1070 Credit: 334,882 RAC: 0 |
Here at CERN we have done quite some investigation into benchmarking VMs in the context of cloud procurement. We do have a benchmarking VM and I could see if we could use it here. The best thing would be to create a benchmarking app. EDIT: Here is a public presentation on the topic. |
Send message Joined: 19 Aug 15 Posts: 46 Credit: 3,626,553 RAC: 0 |
Could you run LINPACK, I assume this would be a relevant test in the context of these Jobs |
Send message Joined: 16 Aug 15 Posts: 966 Credit: 1,211,816 RAC: 0 |
Thanks for the reply. I just wanted it to be as close to the "real" thing as possible, including server response and networking. It would also be interesting, what the bare minimum memory is to run a job, even with a lot of swap usage.( swap could be in RAM and shared between jobs, therefore no slowdown) |
Send message Joined: 19 Aug 15 Posts: 46 Credit: 3,626,553 RAC: 0 |
The people that wrote LINPACK developed a more realistic benchmark that it more I/O bound, HPCG benchmark. I was think that if you took an off the self benchmark could save some time for other project work. I agree if the results are meaning less then it not really saved effort. |
Send message Joined: 12 Sep 14 Posts: 1070 Credit: 334,882 RAC: 0 |
Eric McIntosh, who you my know from Sixtrack, was benchmarking mainframes for procurement many years ago and we discussed the topic recently in the context of cloud procurement. One of the things I took away from the discussion was that you have to benchmark with your own application so you are sure that it touches all the same things by the same amounts. The trick is to find a reference workload that is a good representation of what you actually do. As a result we do have a standard workload for benchmarking and a framework to run it. I will try to add it as an application in this project next week. |
Send message Joined: 12 Sep 14 Posts: 1070 Credit: 334,882 RAC: 0 |
As promised, the benchmarking application is now available :) |
©2025 CERN