Difference between revisions of "CyberShake migration to Summit"

From SCECpedia
Jump to navigationJump to search
Line 11: Line 11:
 
*PreAWP: OK
 
*PreAWP: OK
 
*AWP-GPU-SGT: OK
 
*AWP-GPU-SGT: OK
 +
 +
== UCVM ==
 +
 +
Summit and Titan results agree very well, to floating-point error:
 +
<pre>
 +
Average absolute difference: 3e-13
 +
Average absolute percentage difference: 6.748571e-15%%
 +
Largest difference: 0.00, at index 6882517392.
 +
Largest percent difference: 0.00%, at index 2816011138
 +
Absolute percentage difference histogram:
 +
  0.0001  0.0010  0.0100  0.1000  0.3000  1.0000  3.0000  10.0000  100.0000  1000.0000 
 +
17388000000  0  0  0  0  0  0  0  0  0  0 
 +
100.00%  0.00%  0.00%  0.00%  0.00%  0.00%  0.00%  0.00%  0.00%  0.00%  0.00% 
 +
100.00%  100.00%  100.00%  100.00%  100.00%  100.00%  100.00%  100.00%  100.00%  100.00%  100.00% 
 +
</pre>
 +
 +
== Smoothing ==
 +
 +
  
 
== AWP-GPU-SGT ==
 
== AWP-GPU-SGT ==

Revision as of 19:56, 30 May 2019

This page is being used to gather information on the effort to migrate CyberShake from Titan to Summit.

Software Status

Initially, we are running each phase on Summit and comparing the output with that on Titan.

  • PreCVM: OK
  • UCVM:
  • Smoothing:
  • PreSGT: OK
  • PreAWP: OK
  • AWP-GPU-SGT: OK

UCVM

Summit and Titan results agree very well, to floating-point error:

Average absolute difference: 3e-13
Average absolute percentage difference: 6.748571e-15%%
Largest difference: 0.00, at index 6882517392.
Largest percent difference: 0.00%, at index 2816011138
Absolute percentage difference histogram:
   0.0001   0.0010   0.0100   0.1000   0.3000   1.0000   3.0000   10.0000   100.0000   1000.0000   
17388000000   0   0   0   0   0   0   0   0   0   0   
 100.00%   0.00%   0.00%   0.00%   0.00%   0.00%   0.00%   0.00%   0.00%   0.00%   0.00%   
100.00%   100.00%   100.00%   100.00%   100.00%   100.00%   100.00%   100.00%   100.00%   100.00%   100.00%  

Smoothing

AWP-GPU-SGT

We arbitrarily selected site s2257 and ran the X-component (the AWP X-component, which is the RWG Y-component) on Summit, to compare to the results on Titan.

Numerical comparison

When comparing all SGT points (test-reference), we get:

Average diff = 1.317764e-12, average percent diff = -3.324251%
Largest diff of 0.000150 at float index 1718660606.
Largest percent diff of 5017404928.000000% at float index 8636382051.

When we only consider points greater than 1e-10 (peak SGT values are usually 1e-4 to 1e-3), we get:

Average diff = 1.317764e-12, average percent diff = 0.133713%, average absolute percent diff = 12.979661%
Largest diff of 0.000150 at float index 1718660606.
Largest percent diff of 7148245.500000% at float index 1626932526.

The average percent diff is much less than before, but we included the absolute percent diff, which reveals that many points differ by considerable amounts and there was some canceling going on when looking at the average percent diff.

As a result of this, we decided to more closely investigate what is causing the differences between the two systems.

Initial plots

We identified point ID 71610 as the point which contains the largest difference between the test and reference SGTs.

Below are plots of this point, comparing the Titan and Summit results and the difference. The zoomed-in version focuses on the 25 seconds with the largest differences.

Full 200-second SGT
125-150 seconds

CUDA versions

The default version of CUDA on Titan is 9.1.85 and on Summit it's 9.2.148. We tried rebuilding the AWP code on Summit with 9.1.85 and GCC and rerunning. It didn't make much difference:

Average diff = 1.272838e-12, average percent diff = 0.133971%, average absolute percent diff = 12.971045%
Largest diff of 0.000150 at float index 1718660606.
Largest percent diff of 7148275.000000% at float index 1626932526

X and Y-only work distribution

We tried distributing the work only across X (PX=120, PY=1, PZ=1)but got similar differences.

X:

Average diff = 1.317764e-12, average percent diff = 0.133713%, average absolute percent diff = 12.979661%
Largest diff of 0.000150 at float index 1718660606.
Largest percent diff of 7148245.500000% at float index 1626932526.

We also tried it across Y (PX=1, PY=210, PZ=1). This resulted in a memory error:

*** Error in `/gpfs/alpine/proj-shared/geo112/CyberShake/software/AWP-GPU-SGT/bin/pmcl3d': free(): invalid size: 
0x0000000040728bc0 ***
======= Backtrace: =========
/lib64/libc.so.6(cfree+0x4a0)[0x2000004f9cd0]
/gpfs/alpine/proj-shared/geo112/CyberShake/software/AWP-GPU-SGT/bin/pmcl3d[0x10008d34]
/gpfs/alpine/proj-shared/geo112/CyberShake/software/AWP-GPU-SGT/bin/pmcl3d[0x10008774]
/lib64/libc.so.6(+0x25100)[0x200000485100]
/lib64/libc.so.6(__libc_start_main+0xc4)[0x2000004852f4]

Smaller SGTs

We tried generating SGTs for the SMALL site, which are about 6% the size of s2257, on both Titan and Summit. We actually saw even larger average differences:

Average diff = 6.027801e-12, average percent diff = -1.673267%, average absolute percent diff = 21.367271%
Largest diff of 0.000067 at float index 175412388.
Largest percent diff of 2091817.000000% at float index 1363132518.

Frequency content

We did a frequency content analysis on the largest difference SGT from the SMALL site:

Pt7038 freq comparison.png

Ossian's test

Ossian O'Reilly created a simple, single-GPU test here on Github.

Running his test on Titan and Summit, and running his comparison script, I obtained the following agreement:

Difference (titan, x=1, y=1) (summit, x=1, y=1),  l2 (abs): 2.050260e-04 l2 (rel) 2.090467e-06 l1 (abs) 1.097141e-01 l1 (rel) 7.702747e-06    linf (abs) 6.659655e-06 linf (rel) 1.872744e-06 

Comparing the same two outputs using my scripts, using a cutoff for percent comparisons of 1e-6:

Average diff = 6.023823e-10, average percent diff = -0.026654%, average absolute percent diff = 0.452920%
Largest diff of 0.000007 at float index 128648.
Largest percent diff of 69.524002% at float index 3874621.
Absolute percentage difference histogram:
     0.0001   0.0010   0.0100   0.1000   0.3000   1.0000   3.0000    10.00    100.00    1000.0 
138147   187512   179189   202769   122466   110331    57253    27862     7452        0        0   
 13.37%   18.15%   17.35%   19.63%   11.86%   10.68%    5.54%    2.70%    0.72%    0.00%    0.00%   
 13.37%   31.53%   48.87%   68.50%   80.36%   91.04%   96.58%   99.28%  100.00%  100.00   100.00%

Approximately half the points have a percent difference less than 0.01%, 90% less than 0.3%, and 99% less than 3%.

These two comparisons show us different information: point-to-point comparison can be biased by small values which have little impact on the result. For example, ~80% of the points have values less than 1e-6, but the largest values are greater than 1. L1 comparison can be biased by the largest points, which carry much more weight than points 1 or 2 orders of magnitude lower, which are still important for agreement.

I tried running my code vs Ossian's code on Summit. The difference is how y_rank_F and y_rank_B are handled in Ossian's version. We see differences similar to those above:

Average diff = 3.603304e-09, average percent diff = -0.000870%, average absolute percent diff = 0.585565%
Largest diff of 0.000475 at float index 3272633.
Largest percent diff of 16670.455078% at float index 3272726.
Absolute percentage difference histogram:
  0.0001   0.0010   0.0100   0.1000   0.3000   1.0000   3.0000   10.0000   100.0000   1000.0000   
67238   114994   199943   268809   145799   121488   68547   38022     7930         84           1   
6.51%   11.13%   19.36%   26.03%   14.12%   11.76%    6.64%   3.68%    0.77%      0.01%       0.00%   
6.51%   17.64%   37.00%   63.03%   77.14%   88.91%   95.54%  99.22%   99.99%    100.00%     100.00%

Additional tests

Comparison using Ossian's parameters with CyberShake code on Titan vs Summit, no optimization:

Average diff = 3.000922e-09, average percent diff = 0.013876%, average absolute percent diff = 0.650546%
Largest diff of 0.000475 at float index 3272633.
Largest percent diff of 16670.525391% at float index 3272726.
Absolute percentage difference histogram:
  0.0001   0.0010   0.0100   0.1000   0.3000   1.0000   3.0000   10.0000   100.0000   1000.0000   
72818   114632   210094   283026   131370    108666   63794  36609    11761        84           1   
 7.05%   11.10%   20.34%   27.40%   12.72%   10.52%   6.18%   3.54%    1.14%     0.01%       0.00%   
 7.05%   18.15%   38.49%   65.89%   78.61%   89.13%  95.31%  98.85%   99.99%   100.00%     100.00% 

Bug fix

Ossian identified a bug in most versions of AWP, including the SGT code. The bug is in pmcl3d.c, in lines 267 and 268:

   err       = MPI_Cart_shift(MC1, 0,  1,  &x_rank_L, &x_rank_R );
   err       = MPI_Cart_shift(MC1, 1,  1,  &y_rank_F, &y_rank_B );

If the process calling these functions is a boundary process - that is, it has no previous neighbor in the Y-direction, for example - the neighbor variable (y_rank_F, y_rank_B, x_rank_L, x_rank_R) is set to MPI_UNDEFINED. On Titan, MPI_UNDEFINED is -1, but on Summit it is only required to be less than 0.

Later, the kernel to perform data exchange is called if the neighbor variable is not -1. On Titan, this means the kernel is not called when there is no neighbor, but on Summit it is. Additionally, the buffer used in this exchange is uninitialized, so the effect on Summit is that random values are exchanged with the boundary process, leading to errors.

To fix this, Ossian included the following lines of code in pmcl3d.c, line 273:

   // If any neighboring rank is out of bounds, then MPI_Cart_shift sets the
   // destination argument to a negative number. We use the convention that -1 
   // denotes ranks out of bounds.
   if (x_rank_L < 0) {
           x_rank_L = -1;
   }
   if (x_rank_R < 0 ) {
           x_rank_R = -1;
   }
   if (y_rank_F < 0) {
           y_rank_F = -1;
   }
   if (y_rank_B < 0) {
           y_rank_B = -1;
   }

This ensures the kernel is not called if there is no neighbor. Additionally, all device memory is now initialized to 0.

Verification tests

Following the implementation of Ossian's bug fix, the results on Titan and Summit came into much better agreement.

For a full-scale SGT calculated at site s2257, we see the following differences (minimum differences at 1e-6):

Average diff = -7.335352e-13, average percent diff = 0.000036%, average absolute percent diff = 0.038536%
Largest diff of 0.000006 at float index 240929422.
Largest percent diff of 338.659698% at float index 41967549362.
Absolute percentage difference histogram:
   0.0001   0.0010   0.0100   0.1000   0.3000   1.0000   3.0000   10.0000   100.0000   1000.0000   
478231446   1819753594   2113474276   1241972140   295957433   157813438   27593530   2496656   151869   503   0   
 7.79%   29.65%   34.44%   20.24%   4.82%   2.57%   0.45%   0.04%   0.00%   0.00%   0.00%   
7.79%   37.44%   71.88%   92.11%   96.94%   99.51%   99.96%   100.00%   100.00%   100.00%   100.00% 

Repeating with a cutoff of 1e-10:

19702828989 points used for percent comparison, of 43417008000 total.
Average diff = -7.335352e-13, average percent diff = 0.000288%, average absolute percent diff = 0.491180%
Largest diff of 0.000006 at float index 240929422.
Largest percent diff of 538614.875000% at float index 42529049490.
Absolute percentage difference histogram:
   0.0001   0.0010   0.0100   0.1000   0.3000   1.0000   3.0000   10.0000   100.0000   1000.0000   
559826373   2204732859   3735746248   6725999521   3262492383   2066082069   753902072   280090672   106501047   7143593   312152   
 2.84%   11.19%   18.96%   34.14%   16.56%   10.49%   3.83%   1.42%   0.54%   0.04%   0.00%   
2.84%   14.03%   32.99%   67.13%   83.69%   94.17%   98.00%   99.42%   99.96%   100.00%   100.00%   

We also compared Titan before the bugfix to Titan after (cutoff of 1e-10):