Performance Test

From Yade

Revision as of 13:29, 9 April 2014 by Thoeni (talk | contribs) (→‎Test 2)

Introduction

This page summarises some results for the performance test of YADE (yade --performance) on multicore machines. It should give an idea on how good YADE scales.

Test 1

Two versions of YADE are compared to each other and two different machines are used. The test was conducted on the computing grid of the University of Newcastle by Klaus.

YADE versions:

Machines:

  • AMD: AMD Opteron(tm) Processor 6282 SE (64 cores)
  • Intel: Intel(R) Xeon(R) CPU E5-2670 0 @ 2.60GHz (16 cores)

System:

  • Red Hat Enterprise Linux Server release 6.4 (Santiago)
  • gcc version 4.4.7 20120313 (Red Hat 4.4.7-3) (GCC)

Other:

  • Number of tests per data point in plots: 10

Performance of Parallel Collider

Fig. 1.1: Scaling on Intel machine (left) and AMD machine (right.)

Fig. 1.1 shows how version2 (parallel collider) scales in relation to version1 (NOTE: version1/version2 indicates how much faster the parallel collider is). It is interesting to note that for simulations with less than 100000 particles the scaling is almost not depending on the number of threads and scaling is slightly bigger than one only. For simulations with more than 100000 particles things are looking differently.

Using the total number of cores on the machine is not recommended, e.g. -j12 (and probably -j14) scales better than -j16 on the Intel and -j24 scales better than -j32 on the AMD.

Fig. 1.2: Number of particle vs. ierations per seconds for Intel
Fig. 1.3:Number of particle vs. ierations per seconds for AMD

Figs. 1.2-1.3 show the absolute speed in iterations per seconds for both versions on both machines. It can clearly be seen that the lines for the parallel collider on the right are almost shifted parallel whereas the lines for version1 converge at the 500000 particle point. This means that the parallel collider does not just allow for better scaling (i.e. faster calculation) but as well for more particle to be used in a simulation.


Sample output from Timing

Sample output for timing for -j12 on Intel:

version1 version2 (parallel collider)
Table v1 gpu.png
Table v2 gpu.png

Comparison AMD/Intel

Fig. 1.4: Comparison of Intel vs. AMD for both versions

Fig. 1.4 shows the difference between running version1 and version2 on an Intel or AMD machine. The AMD is generally slower (Intel/AMD>1).

Conclusions

The new parallel collider scales good for the --performance test with more than 100000 particles. The scaling for 500000 particles is really good, i.e. -j12 scales by a factor of 6 for both machines. Intel machines perform better (similar observations have been made here [1]). Finally, I would say that there is an optimum number of threads you should use per simulation. Many cores doesn't always mean much faster. So use your resources wisely.

Finally, it should be mentioned that the results for 500000 particles are extrapolated from 10 iterations since this is the value in the current script. This is far too optimistic. I will post more realistic results soon.

Test 2

Same as Test 1 but with more iterations (1x, 3x and 12x the number of iterations specified in checkPerf.py) and more particles (up to 10^6). However, 3 simulations per data point only. A summary of the full series of results can be downloaded here [2]. All relevant data files and timing stats can be downloaded here [3].

Performance of Parallel Collider

The following figures show the cundall number for both versions. It can be seen that the estimated scaling gets worst when increasing the number of iterations. Nevertheless, the graphs show a more stable Cundall number for the new implementation with the parllel collider.

Fig. 2.1: Cundall number for 1x number of base iterations
Fig. 2.2: Cundall number for 3x number of base iterations
Fig. 2.3: Cundall number for 12x number of base iterations

Some output from Timing

Timing output for -j8 on Intel for various iterations (1x, 3x and 12x the number of iterations specified in checkPerf.py):

iter version1 version2 (parallel collider)
1x
Table v1 iter1.png
Table v2 iter1.png
3x
Table v1 iter3.png
Table v2 iter3.png
12x
Table v1 iter12.png
Table v2 iter12.png

Conclusions

Results indicate again that there is an optimum number of threads. It might be best to try a couple of configurations first since it will certainly be problem dependent.

The timing stats suggest that InteractionLoop and NewtonIntegrator are becoming the bottle neck now.

Test 3

TODO some more real example, feel free to add something...