Host Load Trace Replay

1
Host Load Trace Replay

• Peter A. Dinda
• Thesis Seminar
• 11/23/98

2
Outline

• Motivation for load trace playback
• Interaction with external load
• Example system
• Feedback issue

3
Load Traces for BenchmarkingDistributed Systems

• Real
• Comparable and reproducible
• Think of a set of traces as analogous to a SPEC
  benchmark
• Usable in simulation and experimentation
• Currently synthetic benchmarks
• Poissony process arrivals, power-law process
  run-times, multiplexing
• LTI characterizations (impulse responses, etc)
• Few comparisons

4
Interaction with External Load
External load modulates applied load Other
interactions are possible - human driven apps
5
Load Measurement
unknown sample rate f2 Hz estimated
Trace File
Sample1
h
Sample2
exponential average, tau5 s
f1 Hz

Ready Queue
Kernel
User
6
Load Trace Replay
Load Measure
applied load
Load Generator
Trace File
h-1

measured load
error
-
Sample1
h
Sample2
1.5 load for 1 second gt 100 busy, 50 busy
processes x busy process alternates between
random duration compute and sleep phases so that
x of time is spent in compute phases on unloaded
machine
7
Example One Hour Trace
8
Applied Load
9
Measured Load
10
Error
11
Combined View
12
Distribution of Errors
13
Closer Look
14
Feedback?
Use error signal to better track the load trace
15
The Problem With Feedback
External Load
Applied Load
Load Generator
Load Measure
Trace File

h-1
x
level
Effect of Combined Load
z
error
-
h-1
Feedback would try to make SUM of applied load
and external load in system track the load trace
16
Making Feedback Work
External Load
Applied Load
Load Generator
Load Measure
Trace File

h-1
x
level
Effect of Combined Load
Estimated Effect of External Load
z
error
Signal Separation
-
h-1
Estimated Effect of Applied Load
Load Source Models
17
Conclusions

• Developed reasonable tool for load trace playback
• Identified feedback issue
• Thinking about signal separation