Transcript powerpoint - University of Houston

Characterizing NAS Benchmark Performance
on Shared Heterogeneous Networks
Jaspal Subhlok
Shreenivasa Venkataramaiah
Amitoj Singh
University of Houston
Heterogeneous Computing Workshop, April 15, 2002
Rice01, slide 1
Mapping/Adapting Distributed
Applications on Networks
Data
Model
Sim 2
Pre
Stream
Vis
Sim 1
Application
?
Network
Rice01, slide 2
Automatic node selection
Select 4 nodes for execution : Choice is easy
m-7
m-6
Congested
route
Busy
nodes
m-5
m-8
m-4
selected
nodes
Compute nodes
Routers
m-1
m-2
m-3
Rice01, slide 3
Automatic node selection
Select 5 nodes: choice depends on application
m-7
m-6
Congested
route
Busy
nodes
m-5
m-8
m-4
selected
nodes
Compute nodes
Routers
m-1
m-2
m-3
Rice01, slide 4
Mapping/Adapting Distributed
Applications on Networks
Data
Model
Sim 2
Pre
Stream
Vis
Sim 1
Application
?
Network
1) Discover application characteristics and model
performance in a shared heterogeneous environment
2) Discover network structure and available resources
(e.g., NWS, REMOS)
3) Algorithms to map/remap applications to networks
Rice01, slide 5
Methodology for Building Application
Performance Signature
Performance signature = model to predict application
execution time under given network conditions
1. Execute the application on a controlled testbed
2. Measure system level activity during execution
– such as CPU, communication and memory usage
3. Analyze and discover program level activity
(message sizes, sequences, synchronization waits)
4. Develop a performance signature
•
No access to source code/libraries assumed
Rice01, slide 6
Discovering application characteristics
Executable
Application
Code
ethernet switch
(crossbar)
100 Mbps
links
500MHz
Pentium Duos
Benchmarking
on a controlled
testbed and
analysis
Model as a
Performance
Signature
• capture patterns of
CPU loads and traffic
during execution
Rice01, slide 7
Results in this paper
Executable
Application
Code
ethernet switch
(crossbar)
100 Mbps
links
500MHz
Pentium Duos
Benchmarking
on a controlled
testbed
Measure
performance
with resource
sharing
• capture patterns of
CPU loads and traffic
during execution
Demonstrate that measured resource usage on a testbed is
a good predictor of performance on a shared network for
NAS benchmarks
Rice01, slide 8
Experiment Procedure
• Resource utilization of NAS benchmarks measured
on a dedicated testbed
– CPU probes based on “top” and “vmstat” utility
– Bandwidth using “iptraf”, “tcpdump”, SNMP queries
• Performance of NAS benchmark measured with
competing loads and limited bandwidth
– Employ dummynet and NISTnet to limit bandwidth
• All measurements presented are on 500MHz Pentium
Duos, 100 Mbps network, TCP/IP, FreeBSD
• All results on Class A, MPI, NAS Benchmarks
Rice01, slide 9
Discovered Communication Structure of NAS
Benchmarks
0
1
0
1
0
1
2
3
2
3
2
3
BT
CG
IS
0
1
0
1
0
1
2
3
2
3
2
3
LU
MG
SP
0
1
2
3
EP
Rice01, slide 10
Percentage increase in execution time
Performance with competing
computation loads
140
120
All nodes are loaded
Most busy node loaded
Least busy node loaded
100
• Increase beyond 50% due
to lack of coordinated
(gang) scheduling and
synchronization
• Correlation between low
CPU utilization and smaller
increase in execution time
(e.g. MG shows only ~60%
CPU utilization)
80
60
40
20
0
EP BT CG IS LU MG SP
• Execution time is lower if
least busy node has a
competing load (20%
difference in the busyness
level for CG)
Rice01, slide 11
140
16
120
14
12
100
10
80
8
60
6
40
4
20
2
0
0
CG
IS
MG
SP
BT
LU
Link network traffic (Mbps)
Percentage increase in
execution time
Performance with Limited Bandwidth
(reduced from 100 to 10Mbps) on one link
EP
Close correlation between link utilization and
performance with a shared or slow link
Rice01, slide 12
500
80
450
70
400
60
350
50
300
250
40
200
30
150
20
100
10
50
0
0
IS
CG
SP
MG
BT
LU
Total network traffic (Mbps)
Percentage increase in
execution time
Performance with Limited Bandwidth
(reduced from 100 to 10 Mbps) on all links
EP
Close correlation between total network traffic and
performance with all shared or slow links
Rice01, slide 13
Results and Conclusions
(not the last slide)
• Computation and communication patterns can be
captured by passive, near non-intrusive, monitoring
• Benchmarked resource usage pattern is a strong
indicator of performance with sharing
– strong correlation between application traffic and
performance with low bandwidth links
– CPU utilization during normal execution a good
indicator of performance with node sharing
Synchronization and timing effects were not
dominant for NAS Benchnmarks
Rice01, slide 14
Discussion and Ongoing Work
(the last slide)
• Capture application level data exchange pattern from
network probes (e.g. MPI message sequence, sizes)
– slowdown different for different message sizes
• Infer the main synchronization/waiting patterns
– Impact of unbalanced execution and lack of gang
scheduling
• Capture impact of CPU scheduling policy for
accurate prediction with sharing
– Policies try to compensate for waits
Goal is to build a quantitative “performance signature”
to estimate execution time under any given network
conditions, and use it in a resource management
prototype system
Rice01, slide 15