The scaling is looking awesome! Obviously the MC sims will need work and other tests will be required. But the functionality of the branch is looking very promising and exciting. I have more responses below. Please expect to have delayed responses to the previous messages. I will respond to your posts Gary, but I'm three days away from leaving Australia and am flat out organising and packing. I'll then be spending a week in London before heading to Germany. It could be a few weeks before I'll be able to properly response to posts. On 4/20/07, Gary S. Thompson <garyt@xxxxxxxxxxxxxxx> wrote:
Dear All I have now had a chance to do some true multi tasking on our local cluster with real overhead from intreprocess communication and the results are as follows processors min eff mc eff grid eff 1 18 100 80 100 134 100 2 9 100 4 5 90 8 3 75 16 1 112.5 32 1 56.25 8 31.25 4 104.6 and the picture that speaks 1000 words key top graph black line achieved runtimes top graph red line expected runtimes with perfect scaling efficency bottom graph scaling efficiency some notes 0. data was collected on one of chris's small data sets containing 28 residues not all of which are active for minimisation columns processors - no slave mpi processors min - time for a minimisation of models m1-m9 with a fixed diffusion tensor eff - approximate parallel efficiency expected runtime/ actual runtime
It would be interesting to see if the efficiencies all converge to 100% when a larger number of spin systems are minimised. Maybe duplicating the data a number of times creating an artificially large protein would be useful in that regard.
mc - 256 monte carlo calculations eff - efficiency of the above grid - a grid search on a anisotropic diffusion tensor 6 steps
Do you mean the spheroid (axially symmetric) or the ellipsoid, as both are anisotropic? I would recommend increasing the number of steps in this grid search if MPI is running. With that type of scaling efficiency, I would recommend 11 or 21 increments per dimension on a 32 processor cluster. A drop from 134 min to 4 min is huge!
eff - efficency of the above tests were run on a cluster of opterons using gigabit ethernet and mpi 1. these results are crude wall times as measured by pythons time.time function for the master but they do not include startup and shutdown overhead
They should be more than accurate enough for these types of comparison.
2. these tests are single point measurements there are no statistics
For now statistics are unnecessary.
3. timings were rounded to 1 second, so for example we must consider data points for more than 16 processors for the min run to be suspect
That would explain the 56% efficiency for 32 processors.
The results also highlight up some interesting considerations 1 our local cluster has very poor disk io, with the result that when i first ran the calculations I saw no multiprocessor imporvements on the min run (in actual fact it got worse!) I got round this for this crude test by switching off virtually all text output from the various minimisation commands. Now obviously this isn't a long term solution but I can thing of other methods e.g using an output thread thread on the master or output batching that would improve these results.
Both options sound good. This type of threading is not very complicated, although debugging blocked threads is hell. Sending the minimisation print out in one hit at the end would be very useful as well.
2. comparison of the results from the grid calculation and the other calculations are quite informative. clearly the grid results are excellent. I believe this is because I am returning individual subtask results to the master as they complete and the resulting overhead due to waiting for the master is a problem. To make this (clearer?) here is an example: in the case of the mc run I will take the 256 mc runs and distribute a batch of 8 to each processor (in the case of a 32 processor run) I then resturn the results individually as they complete I believe this can lead to access to the master being the bottleneck (this is most probably due to output ovrehead on stdout again, though problems with contention due to coherence of the calculation length could also be a problem ).
In the Monte Carlo simulations, all of the output of the minimisations is suppressed. Therefore the sending of the minimisation print out shouldn't be the issue as nothing needs to be sent. There must be something else at play. Finding out what this is exactly is important before an investment into the threading or batching is made.
In the case of the grid there are no subtasks as the grid is almost ideally sub divided by processor so only one task is run on each slave.
Do the MC sims have the same scaling efficiency if only one simulation is sent to each processor at once? Does the efficiency increase or decrease?
I can see at least two answers to this. One is to batch the return of results so all results get returned at once and the second is to have an output thread on the master separate from the thread ervicing mpi calls so processing of returned data doesn't block the master and thus the rest of the cluster.
As I mentioned above, both would be useful. However it would be good to know which will be the most beneficial for increasing efficiency before implementation. It could be that the one or the other will not result in any significant improvements.
I have some more comments to follow on the design of the current minimisation interface, how text output from the commands is controlled, and unit testing but these will have to follow in another message later on regards gary n.b. if the pciture doesn't dsiplay well my apolergi
The picture at https://mail.gna.org/public/relax-devel/2007-04/msg00048.html displays perfectly. I've been thinking about how you could release an MPI relax version prior to the merging of a patch into the 1.3 line. You could release a relax-1.3.0-gt version (gt for Gary Thompson). This could itself have a few versions associated with it (I don't know how they would be called though). What do you think Gary? Cheers, Edward