Workpackage 2 Proposal for an Extension of ASL PowerPoint PPT Presentation

presentation player overlay
1 / 10
About This Presentation
Transcript and Presenter's Notes

Title: Workpackage 2 Proposal for an Extension of ASL


1
Workpackage 2Proposal for an Extension of ASL
  • T. Fahringer, M. Gerndt, J. Nabrzyskin, B.
    Parady, O. Plata, D. Tessera
  • Based on work done at WP2 Poznan meeting in
    April.

2
Outline
  • Extension of ASL
  • Property Templates
  • Meta-properties
  • Future Work

3
Apart Specification Language (ASL)
  • ASL supports description of
  • performance related data
  • performance properties.
  • Several drawbacks of ASL
  • Many properties are similar, yet must be
    specified separately. ? introduce property
    templates
  • New properties cannot easily be described by
    existing ones. ? introduce meta-properties

4
Property Templates
  • Many performance properties differ only in
    accessed performance data.
  • E.g. communication, I/O, and synchronization
    costs.

Property communication_costs (MPIRegion r,
Experiment e, Region rank_basis) LET float
cost summary(r,e).sums.comm_time IN CONDITION
cost gt 0 CONFIDENCE 1 SEVERITY
cost/duration(rank_basis,e)
Property io_costs (MPIRegion r, Experiment e,
Region rank_basis) LET float cost
summary(r,e).sums.io_time IN CONDITION cost gt
0 CONFIDENCE 1 SEVERITY cost/duration(rank_b
asis,e)
Property templates could be used for a generic
description of similar performance properties
5
Property Templates (cont.)
PROPERTY TEMPLATE cost ltfloat cost_func(RegionSumm
ary)gt (Region r,
Experiment e, Region RankBasis)
LET costcost_func(summary(r,e)) IN CONDITION
cost gt 0 CONFIDENCE 1 SEVERITY
cost/duration(RankBasis,e)
float io_cost_func(RegionSummary
rs)rs.sums.io_time float sync_cost_func(RegionSu
mmary rs)rs.sums.sync_time float
comm_cost_func(RegionSummary rs)rs.sums.comm_time

Define new properties based on template by using
a defined function as template parameter.
PROPERTY costltio_cost_funcgt io_cost PROPERTY
costltsync_cost_funcgt sync_cost PROPERTY
costltcomm_cost_funcgt comm_cost
6
Meta-properties (1)
  • Meta-properties can be used to define new
    performance properties based on existing ones.
  • Meta-properties are again defined as property
    templates.
  • Meta-property parameters can be
  • functions with defined signature
  • performance properties with defined context
    signature

PROPERTY TEMPLATE property_on_all_processes
ltPROPERTY x(Region,Experiment,Process,Region)gt
(Region r, Experiment e, Region RankBasis)
CONDITION Forall p IN e.processors SUCH THAT
condition(x(r,e,p,RankBasis)) AND
confidence(x(r,e,p,RankBasis))1 CONFIDENCE
1 SEVERITY max(severity(x(r,e,p,RankBasis))
WHERE p in e.processes)

7
Meta-properties (2)
PROPERTY TEMPLATE cost_per_process
ltFLOAT cost_func(RegionSummary)gt (Region r,
Experiment e, Process p, Region RankBasis)
LET cost cost_func(summary(r,e,p)) IN CONDIT
ION cost gt 0 CONFIDENCE 1 SEVERITY
cost/duration(RankBasis,e)
Float io_cost_func(RegionSummary
rs)rs.sums.io_time Float sync_cost_func(RegionSu
mmary rs)rs.sums.sync_time Float
comm_cost_func(RegionSummary rs)rs.sums.comm_time

PROPERTY cost_per_processltio_cost_funcgt
io_cost_per_process PROPERTY cost_per_processltsyn
c_cost_funcgt sync_cost_per_process PROPERTY
cost_per_processltcomm_cost_funcgt
comm_cost_per_process
8
Meta-properties (3)
  • PROPERTY TEMPLATE property_on_all_processes
  • ltPROPERTY x(Region,Experiment,Process,Region)gt
  • (Region r, Experiment e, Region RankBasis)
  • CONDITION Forall p IN e.processors SUCH THAT
  • condition(x(r,e,p,RankBasis)) AND
    confidence(x(r,e,p,RankBasis))1
  • CONFIDENCE 1
  • SEVERITY max(severity(x(r,e,p,RankBasis)) WHERE
    p in e.processes)

PROPERTY property_on_all_processes
ltio_cost_per_processgt io_cost PROPERTY
property_on_all_processes ltsync_cost_per_processgt
sync_cost PROPERTY property_on_all_processes
ltcomm_cost_per_processgt comm_cost
9
Future Work
  • Extend ASL Syntax for meta-properties and
    templates
  • Case-studies of ASL for
  • a few machines (e.g. Cray T3E)
  • Case studies for OpenMP
  • Formalize search for bottlenecks based on ASL.

10
CASE-Studies
  • WP2
  • Specify properties for Cray T3E
  • Specification of search process for properties
  • WP3
  • Experiment planing
  • Effect of multiple sources of performance data.
  • Set of test kernels/applications
Write a Comment
User Comments (0)
About PowerShow.com