DisCostiC v2.0.0
Distributed Cost in Cluster: a cross-architecture simulator for massively parallel MPI programs
Loading...
Searching...
No Matches
📶 DisCostiC output

Part of the main README.

🌐 1. Web interface with Google Chrome browser (DMS file format) 💬

Upon completion of the run, the controller rank merges the per-rank Chrome traces into <chromevizfilename>.dms. The prefix comes from chromevizfilename in config.cfg (the shipped file uses DisCosTiC). The .dms file is a Chrome trace JSON document. Load it in Google Chrome with:

chrome://tracing

📊 2. Graphical user interface with Vampir (OTF2 file format) 💬

make otf2 runs python visualization/chrome2otf2.py -i DisCosTiC.dms -o DisCosTiC -c. That writes an OTF2 archive in the DisCosTiC directory (DisCosTiC/traces.otf2). Open a small trace in Vampir with:

vampir DisCosTiC/traces.otf2

📊 3. Graphical user interface with ITAC (STF file format) 💬

Two different ITAC products exist.

The modeled application is traced inside the simulator when USE_ITACVIZ is enabled. Rank files are merged to DisCosTiC.avt. make stf runs visualization/convert_itac_stf.sh on that file and writes DisCosTiC.single.stf.

Separately, make trace_MPI or make trace_all rebuilds the simulator with Intel -trace so ITAC can record the simulator process itself. There is no separate launch script for that trace. Open the resulting STF file in Intel Trace Analyzer.

🗄️ 4. Statistical or log data 💭

With Verbose = 1 in config.cfg, DisCostiC prints per-rank runtimes and the computation and communication model parameters. The listing below is an example of that log. The current binary prints Version : v2.0.0 (Second release).

----------------------------------------------------------------
DisCostiC
----------------------------------------------------------------
Full form: Distributed Cost in Cluster
Version: v1.0.0
Timestamp: Tue Jan 20 11:16:42 2024
Author: Ayesha Afzal <ayesha.afzal@fau.de>
Copyright: © 2024 HPC, FAU Erlangen-Nuremberg. All rights reserved
HEAT program in FILE mode of kernel is simulated for 100000 iterations and 72 processes
----------------------------------------------------------------
COMPUTATIONAL PERFORMANCE MODEL
----------------------------------------------------------------
Code simulating on IcelakeSP_Platinum-8360Y chip of cores: {1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 }
Single-precision maximum chip performance [GF/s]: {614.4 , 1228.8 , 1843.2 , 2457.6 , 3072 , 3686.4 , 4300.8 , 4915.2 , 5529.6 , 6144 , 6758.4 , 7372.8 , 7987.2 , 8601.6 , 9216 , 9830.4 , 10444.8 , 11059.2 , 11673.6 , 12288 , 12902.4 , 13516.8 , 14131.2 , 14745.6 , 15360 , 15974.4 , 16588.8 , 17203.2 , 17817.6 , 18432 , 19046.4 , 19660.8 , 20275.2 , 20889.6 , 21504 , 22118.4 }
Double-precision maximum chip performance [GF/s]: {307.2 , 614.4 , 921.6 , 1228.8 , 1536 , 1843.2 , 2150.4 , 2457.6 , 2764.8 , 3072 , 3379.2 , 3686.4 , 3993.6 , 4300.8 , 4608 , 4915.2 , 5222.4 , 5529.6 , 5836.8 , 6144 , 6451.2 , 6758.4 , 7065.6 , 7372.8 , 7680 , 7987.2 , 8294.4 , 8601.6 , 8908.8 , 9216 , 9523.2 , 9830.4 , 10137.6 , 10444.8 , 10752 , 11059.2 }
----------------------------------------------------------------
COMMUNICATION PERFORMANCE MODEL
----------------------------------------------------------------
Code simulating on InfiniBand with IntelMPI library
Hockney model parameters (intra chip): 41185.7 [ns] latency, 0.079065 [s/GB] inverse bandwidth
Hockney model parameters (inter chip): 64567.6 [ns] latency, 0.124122 [s/GB] inverse bandwidth
Hockney model parameters (inter node): 47628.7 [ns] latency, 0.091516 [s/GB] inverse bandwidth
Hockney model parameters (inter cluster): 47628.7 [ns] latency, 0.091516 [s/GB] inverse bandwidth
LogGP model additional parameters (intra chip): 24429.9 [ns] o, 199.008 [ns] g, 0.065 [ns] O
LogGP model additional parameters (inter chip): 39748.3 [ns] o, 338.249 [ns] g, 0.085 [ns] O
LogGP model additional parameters (inter node): 47938.9 [ns] o, 4980.51 [ns] g, 0.084 [ns] O
LogGP model additional parameters (inter cluster): 47938.9 [ns] o, 4980.51 [ns] g, 0.084 [ns] O
Eager threshold: 65535 bytes
----------------------------------------------------------------
Simulation runtime in DisCosTiC: 15.7239 [s]
----------------------------------------------------------------
----------------------------------------------------------------
FULL APPLICATION PERFORMANCE (for all MPI processes)
rank runtime [s]
----------------------------------------------------------------
0 2154.83
1 2154.83
. .......
. .......
. .......
71 2154.67
----------------------------------------------------------------