Skip to content

Latest commit

 

History

History
140 lines (101 loc) · 3.75 KB

README-Install.md

File metadata and controls

140 lines (101 loc) · 3.75 KB
Raw

Prerequisites

Environment

  • GCC
  • Python (required: v2 for perf)
  • Python Module(s)
    • pefile for perf

Benchmarks

  • OpemMP
  • MPI

Building & Installing

  1. Options:

    DEVELOP=1

MG_PERF_CC=<linux-perf-cc>

export MG_XLIB_ROOT=<optional-xlib-build-path>
export MG_PERF_ROOT=<optional-perf-build-path>

  2. Build libraries

    cd <memgaze>/xlib
make xlib
make perf # optional

  3. Build MemGaze

    cd <memgaze>
make PREFIX=<install-prefix> install

Using

MemGaze has four main steps.

  1. Instrument application binary (or libary) <app> for memory tracing. Generate instrumented <app> and auxiliary data within directory <inst-dir>, which defaults to <app-dir>/memgaze-<app>. Instrumented <app> is <inst-dir>/<app>.

    Note: If instrumenting <app>'s libraries, must instrument each library and relink instrumented <app>.

    memgaze-inst [-o <inst-dir>] <app-dir>/<app>

Important contents of <inst-dir>:

  • <app> : Instrumented <app>
  • <app>.binanlys : Static binary analysis
  • <app>.hpcstruct : HPCToolkit structure file

Steps:

  • Read binary (with dyninst)
  • Create CFG and dependence graph for each routine to determine each load's classification (strided, indirect, constant)
  • Selectively insert ptwrite.
  • Generates 1) load classification data and 2) mapping from original IP to instrumented IP (libexec/memgaze-inst-cat)
  • Generates HPCToolkit structure file

Uses:

libexec/memgaze-inst-cat <tracedir>/<app>.binanlys.log <tracedir>/<app>.binanlys0

  1. Trace memory behavior of application <app>, which is normally <inst-dir>/<app>. Generates output in <trace-dir>, which defaults to ./memgaze-<app>-trace-b<bufsz>-p<period>.

    By default collect a sampled trace with period ; if period is 0 then collect a full (non-sampled) trace. Sample event can be either load-based or time-based.

    memgaze-run -p <period or 0> -b <bufsz> [-e pt-load | pt-time | ldlat] [-o <trace-dir>] -- <app> <app-args>

    Important contents of <trace-dir>:

    • <memgaze.config>: Configuration arguments
    • <app>.perf : Tracing data (Linux perf)

  2. Translate tracing and profiling data within <trace-dir> into open formats and place results therein.

    memgaze-xtrace <trace-dir>

New contents of <trace-dir>:

  • <app>.trace : Memory references

  • <app>.callpath : Call paths

  • Extract data file with perf-script

  • Convert IP offsets (from perf script) to full static IPs and combine two-address loads into single trace entry.

  • Remove data collection errors from trace

  • Separate memory references and call paths

Uses:

perf script libexec/perf-script-*
libexec/memgaze-xtrace-normalize <trace-dir>.trace

  1. Analyze memory behavior. Default analyzes memory operations over time using execution interval tree and generates metrics characterizing accesses, spatio-temporal reuse, and access patterns. As inputs, takes <trace-dir>.

    memgaze-analyze [-o <output>] <trace-dir>

Currently, memgaze-analyze's default output includes two parts:

  • execution window histograms
  • function analysis summary (exclusive), averaged over whole trace, for all functions

Note: -f : Focus the above analysis on 'func', i.e., the effective trace includes all accesses between first and last instance of .