Print

Print
That is my theory, but sometimes the sawtooth happened and then the next time, running the same test exactly, it would be fine. The boers that are part of the cluster aren't being used for anything else right?

I'll try reproducing the sawtooth problem and then imposing ForceLocal using an input dataset which is spread out over all the worker boers and see if I can get through many successive test runs without the sawtooth problem appearing. That should better inform us about what is going on. The sawtooth effect appeared for datasets which are more or less evenly spread over all nodes, so I would not blame that problem on the boer0125 deficit that the one dataset has.

I think I did see the network latency with 36 workers, but only for the dataset which was missing files on boer0125, so I don't really consider this a problem.

Yang, Wei wrote:
[log in to unmask]" type="cite"> 
OK, let's postpone it to 7/7 and see if we have enough to make a decision on hardware. I will likely be at CERN for a week from 7/12-16. 

BTW, do you have plans to other measurement, in addition to a simple test on atlint01? It looks like you saw the network latency when you ran 36 workers, though just slightly. Well, the whole idea of proof is to avoid network as much as possible. The sawtooth effect, if according to your explanation that it was because some workers are slow, maybe contributed by the slight uneven data distribution amount cluster nodes, and that boer0125 doesn't have local files at all (because it was down when you copied data?). Is imposing local file accessing only something useful?

TProof *p = TProof::Open("boer0123");
p->SetParameter("PROOF_UseTreeCache", 1);  // this is the default
p->SetParameter("PROOF_ForceLocal", 1);

regards,
Wei Yang  |  [log in to unmask]  |  650-926-3338(O)


On Jun 22, 2010, at 12:06 PM, Bart Butler wrote:

  
I am available, though Ariel is at BOOST this week and may not be able 
to make it.

Yang, Wei wrote:
    
Now that Bart has made lots of measurement,  should we resume our Proof meeting tomorrow? In case you forgot, the meeting is 9am PDT (6pm CERN?), call in number is 510-665-5437 # 3935

regards,
Wei Yang  |  [log in to unmask]  |  650-926-3338(O)