DGX Nightly Benchmark run 20201201

[quasiben]

Historical Throughput

Raw Data

<Client: 'tcp://127.0.0.1:37017' processes=10 threads=10, memory=540.94 GB>
Distributed Version: 2.31.0.dev0+58.g863b0105
simple
5.971e-01 +/- 4.167e-02
shuffle
2.224e+01 +/- 5.193e-01
rand_access
1.039e-02 +/- 1.207e-03
anom_mean
9.748e+01 +/- 1.237e+00

Raw Values

simple
[0.63938522 0.56597996 0.56482959 0.5679853 0.63865495 0.58807492
0.61041498 0.64601135 0.51409435 0.63523746]
shuffle
[21.48743486 22.00875306 21.55290961 21.57217765 22.52186704 22.36727738
22.85407138 22.73326921 22.85354495 22.49844527]
rand_access
[0.01017618 0.00788522 0.01258779 0.00999117 0.01085591 0.01135015
0.00982213 0.01149797 0.00977087 0.0099709 ]
anom_mean
[99.054811 97.62341428 98.87981558 94.83508325 97.62352633 97.95679402
97.43829679 95.69920492 98.11229801 97.60029459]

Dask Profiles

• Shuffle
  Profile
• Random Access
  Profile
• Simple
  Profile
• Anom
  Mean

Scheduler Execution Graph

Client Execution Graph

[jakirkham]

JFYI this is the first profile where we have started compiling bits of Distributed with Cython. Previously we were just running pure Python. Though we are still working on optimizing Distributed. So wouldn't expect this to change a lot.

[jakirkham]

Interestingly this is the first one where shuffle takes a bit less time. In addition to enabling Cythonization, we optimized check_idle_saturated ( dask/distributed#4289 ) a slow function playing a role in the 2 slowest transitions: transition_processing_memory and transition_waiting_processing. Also we annotated ClientState and typed all variables holding ClientState. ( dask/distributed#4290 )