There are two types of benchmarks, one is by single node, the other is by multi nodes. The results up to now are in here.

Notice: su3_rmd_symzk1_asqtad of MILC applications is used in this benchmarks.

• Single node benchmarks
• 1cpu
• SMP
• 2cpu
• Multi node benchmarks
• PBS batch
• Collects TRACE data
• (addition) MPE

This benchmark collects the performance of each problem (lattice) size on one node. Three types are in there, "1cpu", "SMP", and "2cpu".

So you change the directory to QCD-benchmark-1.0/benchmarks/SingleNode . There are four makefiles.

• Makefile
• make.bin
• make.1cpu
• make.smp
• make.2cpu

[Reminder] .SMP-node file is in there. It needs to run make.smp(MPICH). You MUST set your machine-name on .SMP-node. If you forgot to set it, make.smp will not work.

When you want to run benchmarks, you just type "make" here.

1cpu
using One cpu on One node.

SMP
using Two cpu on One node which is SMP.

2cpu
using Two cpu on Two nodes, But they do NOT cooperate, Just concurrent.

Multi node benchmarks

This benchmark collects the performance of each number of CPUs on a fixed lattice size. Typically, you make some batch script and throw them to PBS. I describe this handle on below "PBS batch".

On the other hand, you want to know about timing on the application. Then you use a TRACE benchmark. This detail is in "Collects TRACE data".

In any case, you have to compile applications for this benchmark.

% make

PBS batch

usage: batch4PBS [options] MILC_application

Options are

• -Ethernet/-Myrinet

Choose the network type, Ethernet or Myrinet.
• -SMP

Choose for SMP environment.
• -l#

Set lattice sizes. Its format is like that -l12,16. This means you set two lattice 12x12x12x24 and 16x16x16x32.
• -p#

Set number of processors. Its format is like that -p8,4,2,1. This means you set four processors 8, 4, 2, and 1.
• -rd<directory>

Choose benchmark results stored directory.
• -default

Print default values
• -help

Print usage and options

e.g.)
% ./batch4PBS -Myrinet -SMP -l4 -p8,4,2,1 -rdres bin/su3_rmd_symzk1_asqtad.ch_gm

Collects TRACE data

usage: run_trace [options] MILC_application

Options are

• -n #

Number of processors
• -t <nettype>

Network type (Ethernet of Myrinnet)
• -i <infile>

Input file name
• -o <outfile>

Output file name
• -d <resdir>

Result storage directory
• -s <nodefile>

Nodes list file

e.g.)
% ./run_trace -n 4 -t Myrinet -i samples/in.lat4444 -o out.log -d trace-res -s nodes bin/su3_rmd_symzk1_asqtad.ch_gm-trace

NOTICE: "Nodes list file"'s format is the same of $PBS_NODEFILE. If you use PBS, you do not mention about nodefile. BUT you want to run this scripts on manually, you have to specify your using machine names on a node file and use -s option.

(addition) MPE

Perhaps! you want to get SLOG or CLOG data by MPE. I prepare the MPE target in "make.bin". You will get applications linked MPE libraries as follows:

% make -f make.bin su3_rmd_symzk1_asqtad.ch_p4-trace-mpe
	or
% make -f make.bin su3_rmd_symzk1_asqtad.ch_gm-trace-mpe

Before running those application, you have to set the environment variable MPE_LOG_FORMAT.

% export MPE_LOG_FORMAT=SLOG
	or
% export MPE_LOG_FORMAT=CLOG

Then you just run the application. After running, you will obtain CLOG or SLOG data in the bin directory where the application exists.

---

ToDo

% cd QCD-benchmark-1.0/benchmarks/SingleNode
Set .SMP-node
% make

% cd QCD-benchmark-1.0/benchmarks/MultiNode
% make
% ./batch4PBS -Myrinet -SMP -rdresults bin/su3_rmd_symzk1_asqtad.ch_gm
	or
% ./run_trace -n 4 -t Myrinet -i samples/in.lat4444 -d trace-res -s mynode bin/su3_rmd_symzk1_asqtad.ch_gm-trace

[back]

---

mats@fnal.gov