Dynamic%20Instrumentation%20of%20Large-Scale%20MPI%20and%20OpenMP%20Applications

1
Dynamic Instrumentation of Large-Scale MPI and
OpenMP Applications
Christian Thiffault, Michael Voss, Steven T.
Healey, Seon Wook Kim
Presented By Jaydeep Marathe
2
The Basic Idea ..

• Instrumentation needed to collect collect
  performance data.
• Static (compile or link time) instrumentation
• Dynamic (run-time) instrumentation

• Feasibility study for dynamic instrumentation
  and dynamically
• controlled static instrumentation.

• Paper asserts Dynamic methods better than
  static instrumentation with
• respect to execution overhead trace data size

But thats still debatable, as we shall see ...
3
Dynamic Instrumentation (with DPCL)
Executing Program
JMP _Test
Probe Point
4
Dynamic Control of Static Instrumentation
Instrumentation Library
Compiler-Instrumented Program
Conf_sync() return
...
...
Conf_sync()
...
Read Active List
...
...
Conf_sync()
Point Active ?
...
...
...
Conf_sync()
...
Instrument compile-time, decide activation at
run-time.
5
Instrumentation Tool Dynprof

• For mixed (?) OpenMP MPI programs

GuideTrace OpenMP Library
KAI Guide Compiler
Application Source
Dynprof
VampirTrace Library
Executable
MPI Library

• Applicable only to the KAI OpenMP compiler
• Instrumentation VampirTrace call to log trace

Trace Data
6
Experiments ..

• Compare execution times for static vs. dynamic
  instrumentation.
• 4 ASCI benchmarks - Smg98, Sppm, Sweep3d
  (MPI), Umt98 (OpenMP)
• IBM Power3 144-node system, each node has 8
  Power3 processors.
• Measure execution times with 5 scenarios

• FULL Full Static Instrumentation Trace logging
• FULL-OFF Full Static Instrumentation, no trace
  logging
• SUBSET Full Static Instrumentation, only
  selected functions active.
• None No instrumentation (Baseline)
• Dynamic Dynamically inserted instrumentation.

7
Results Smg98
8
Results Smg98 (contd)
Static approach (SUBSET), more expensive than
dynamic insertion (DYNAMIC).
WHY ?? - No explanation in paper.
9
Results Smg98 ? Another interpretation
Observe values for FULL, FULL-OFF,SUBSET. Very
Close !!
Compiler instrumentation inefficiency ? skews
results ! ? dynamic approach looks better.
10
Time to create and instrument

• Significant cost, but only occurs at startup.
• May significantly perturb application, when
  instrumented during run-time.

11
Summary and Conclusions

• Tool to instrument MPI and OpenMP programs
  (Dynprof)
• Trace logging/visualization through existing
  VampirTrace library.
• Compare costs of static,dynamic and
  dynamically controlled static
• instrumentation.
• Asserts that dynamically inserted
  instrumentation is less expensive, as compared to
  static instrumentation.