1. Communication Pattern Based Node Selection for Shared Networks
Srikanth Goteti
Interactive Data Corp
Jaspal Subhlok
University of Houston
AMS Symposium 2003

2. Resource Selection for Network/Grid Applications
Model
Data
GUI
Sim 1
Pre
Stream
Application ?
where is the best performance
Network

3. Current Approaches to Node Selection
Model
Data
GUI
Sim 1
Pre
Stream
Measure and model network properties, such as available bandwidth and CPU loads (with tools like NWS)
Find “best” nodes for execution based on network status
But expected application performance based on measured network status may not be accurate
depends on application characteristics
translation, e.g., unused bandwidth vs expected throughput
data may be stale as frequent measurements are expensive

4. Performance Skeleton
Performance Skeleton is a synthetic short running program whose execution characteristics mirror the application it represents
An application and its skeleton have similar
communication pattern
synchronization pattern
CPU usage
memory usage
Goal: Performance of a skeleton is directly related to the performance of the application under any condition
e.g., a skeleton executes in .1% of the time the application takes to execute on any part of a shared network

5. Node Selection with Performance Skeletons
Model
Data
GUI
Data
Sim 1
GUI
Model
Pre
Stream
Construct a skeleton for application of interest
Sim 1
Pre
Stream
Select candidate node sets based on network status
Execute the skeleton on them
Select the node set with best skeleton performance to schedule actual application

6. Node Selection Procedure
Construct a performance skeleton
mostly by hand in this paper, subject of ongoing work
Select candidate node sets
identify the communication graph of the application
typically a chain, ring or all-all structure
obtain available bandwidth between nodes with NWS (Network Weather Service) and build a graph
select nodes to “maximize the minimum available bandwidth” between pairs of communicating nodes
best possible node sets based on application structure and network status
Execute the skeleton on each candidate node set
Select the node set with best skeleton performance, map one process to each node

7. Communication Structure of NAS Benchmarks1 1 3 2 CG 1 2 3 2 3 BT IS 1 1 1 2 3 2 3 2 3 LU MG SP 1 2 3 EP

8. 1 2

9. Validation Experiments
Best nodes to execute benchmarks selected by each of the following methods…
skeleton based: full framework discussed
all to all: based on maximizing the minimum available bandwidth between on the network graph
random
…compare performance of the application on nodes selected by each of these procedures on a busy network
Experiments repeated a large number of times to get statistically meaningful results

10. Experimental Framework
Linux cluster of 10 dual CPU 1.7GHz Pentium nodes connected by 100 MHz links and crossbar switch
experiments with Class B NAS MPI benchmark suite
Class W NAS benchmarks (avrg runtime ~1.5 seconds on our cluster) used as skeletons for class B benchmarks
available bandwidth between nodes is varied with Linux iproute2 for the duration of experiments as follows:
path between a pair of nodes is “shared” by S streams
i.e., available bandwidth is set to 1/S of peak
one stream is randomly added to or removed from the cluster every 30 seconds

11. Performance Results: slowdown due to network traffic1 3 2 CG
skeleton based has average slowdown of 20%, versus 40 % for random and 27% for all to all
significant variation across benchmarks, most benefit for CG – it is communication heavy and uses only 3 links

12. Conclusions type slide
Performance skeletons have a role in resource management for grids
removes limitations of using NWS type systems (what you measure versus what you get problem)
A lot more experimentation is needed to establish and validate the concepts
Automatic construction of performance skeletons is a major open challenge
Skeletons may have other uses a fast way of estimating the performance of an application
e.g. on a slow simulated future system