Largest matrix dimension for DSYEV?

[ningfengh]

Currently I ran a benchmark (using the one in the example folder) of the subroutine DSYEV and get the following result.

Initializing CULA...
Initializing MKL...

Benchmarking the following functions:
-------------------------------------
DSYEV
-------------------------------------


-- DSYEV Benchmark --

Size CULA (s) MKL (s) Speedup
------ ---------- ---------- ---------
1024 0.29 0.26 0.8799
2048 0.65 0.45 0.6813
3072 1.55 0.89 0.5721
4096 2.85 3.01 1.0563
5120 4.52 6.47 1.4294
6144 6.84 10.23 1.4943
7168 9.59 14.72 1.5347
8192 13.14 21.90 1.6667
9216 17.24 27.78 1.6116
10240 22.08 36.65 1.6601
11264 27.74 46.85 1.6890
12288 33.88 58.92 1.7391
13312 41.21 73.21 1.7765
14336 49.48 90.24 1.8239
15360 59.19 110.20 1.8618
16384 69.52 156.04 2.2445
17408 80.26 155.48 1.9372
18432 92.58 179.42 1.9381
19456 107.06 206.71 1.9309
20480 121.99 234.07 1.9188
21504 138.50 271.49 1.9602
22528 155.97 308.07 1.9752
23552 175.62 352.68 2.0082
24576


The first thing I saw is that the speedup is not very spectacular. Is it something I would expect from CULA? And then the job crashes at 24576 without any error output. Is this the largest matrix can be managed in our graphical card?

The following are the specs of the system,

dual 8-core 2.4 gigahertz Intel E5-2665 processors
64 gigabytes of memory
two NVIDIA K20 Kepler GPU accelerators

CULA Dense R17
MKL 12.1.1

[john]

CULA requires that all data fit on the GPU simultaneously, so when you hit the limit of the K20, it will be unable to go any further. Note that CULA uses only one GPU for this but MKL uses both processors, so you're comparing $3000 of processors to $3000 of GPUs - 2x is pretty decent, given this setup. (SYEV is algorithmically also not our highest performing routine)

[ningfengh]

[john]

There are no eigensolvers in pCULA, unfortunately.

Memory usage for most CULA routines is only a few megabytes besides what you put into the routine. In the case of SYEV, it's just the size of the matrices.