Wrekavoc a Tool for Emulating Heterogeneity

1
Wrekavoc a Tool for Emulating Heterogeneity

• Louis-Claude Canon Emmanuel Jeannot
• ESEO INRIA-LORIA-ICL
• Angers U. Tennessee
• Knoxville
• HCW 04/25/2006

2
Outline

• On the Importance of Experiments in Computer
  Science
• Wrekavoc a tool for Emulating Heterogeneity
• Design
• Configuration of the nodes
• Experiments

3
Modern Infrastructures are Hard to Model

• Modern infrastructure are complex
• Processors have very nice features
• Cache
• Hyperthreading
• Dual core
• Operating system impacts the performance (process
  scheduling, socket implementation, etc.)
• The runtime environment plays a role
  (MPICH?OPENMPI)
• The same for middleware (Globus?GridSolve)
• Various parallel architectures that can be
• Heterogeneous
• Hierarchical
• Distributed
• Dynamic
• Analytical validation of algorithms is hard and
  sometimes impossible.
• Can we still design and validate algorithms?

4
Experimental Validation

• A good alternative to analytical validation
• Provides a comparison between algorithms
• Provides a validation of the model or helps to
  define the validity domain of the model

5
Methodologies for Doing Experiments
log(cost)
Real systems Real applications In-lab
platforms Synthetic conditions
Real systems Real applications Real
platforms Real conditions
Models Sys, apps, Platforms, conditions
Key system mecas. Algo, app. kernels Virtual
platforms Synthetic conditions
log(realism)
emulation
math
simulation
live systems
6
Wrekavoc a Tool for Emulating Heterogeneity

• We target heterogeneous distributed environment
• Goal experiment distributed algorithm on
  ahomogeneous and centralized cluster
• How transform this cluster into heterogeneous
  environment and control the heterogeneity

7
Goal

• Making a cluster an heterogeneous environment
• CPU speed.
• Mémory.
• Network bandwidth.
• Network latency.
• A real node
• An emulated node.
• Two solutions
• Increase the performance (update the hardware)
• Degrade the performance (by software means)
• Solution 2 Costless and allows for performance
  control.

8
State of the Art
9
A Client-Sever Approach

• The client
• Reads a configuration file,
• Contact each node to be configured.
• One server per node on the cluster
• Run a deamon,
• Configure itself according to client orders,
• Is able to test itself (and send back results to
  the client).

10
Logical Architecture

• The cluster is decomposed into islets.
• 1 islet union of IP addresses intervals
• 152.81.2.12-152.81.2.25-152.81.2.151-152.81.2.1
  76
• Each node of a given islet shares the same
  characteristics
• Network characteristics are define between and
  inside an islet.

Islets of machine sharing Same charactreristics
1
2
3
4
5
6
7
8
Islet sub-network
Logical view
Inter-islet network
Islet sub-network
1
2
3
4
11
Characteristics Definitions

• For each islet one defines
• A seed,
• Internal characteristic of each node of the islet
  (CPU, memory, BW(in/out), latency) according to
• A uniform law inf. bound - up. bound
• A gaussian law avg variance
• Between each islet one defines
• A seed
• BW between islet 1 and 2.
• BW between islet 2 and 1.
• Latency between islets.

12
Example of Configuration File

• islet1 152.81.15.207-152.81.15.209-
  152.81.15.123-152.81.15.254
• SEED 123
• CPU 1000-1200
• BPOUT 10000
• BPIN 10000
• LAT 15-15
• USER ME
• MEM 800000
• islet2 152.81.3.100-152.81.3.100
• SEED -1
• CPU 100-300
• BPOUT 120-185
• BPIN 12-18
• LAT 201
• USER OTHER
• MEM 80000-100000
• !INTER islet1islet2 5-5 100-100 10 -1

• Units
• CPU MHz
• BW KB/s
• Latencey ms
• Mémory Ko

13
Tools for Configuring Nodes

• We want to degrade
• CPU speed
• Allocatable Memory
• Network bandwidth
• Network lantency

14
CPU Speed Degradation

• 3 approaches
• CPU-freq (Linux kernel module that change the CPU
  frequency)
• Advantage very precise.
• Drawback Requires ACPI enabled CPUfew usable
  frequencies (coarse management).
• CPU-burning (A process take some CPU cycles)
• Advantage works on any architecture fine
  management
• Drawback calibrating is hard, degrades net.
  perf. to the same proportion
• CPU-scheduling (a user level scheduler suspend or
  active process execution according to the desired
  degradation).
• Advantage very precise (default method)
• Drawback uses /proc (not portable)

15
Memory Degradation

• Use PAM-limit module of the kernel
• Limit the maximum amount of memory allocatable by
  a malloc.

16
Network Management

• We use (Traffic Control) of iproute2
• Limit ingoing and outgoing bandwidth
• Limit latency (ver. 2.6.8.1 or better).
• Traffic control depends on IP addresses

17
Experiments

• Several experiments on GdX
• Configuration time
• Micro-benchmark
• Impact of CPU degradation against available
  bandwidth
• Algorithms of the literature

18
Configuration Time
19
Micro-benchmark
20
CPU Degradation vs. Bandwidth Degradation
21
Matrix Mutliply Algorithms of the Literature
20 machines 5 at 100 5 at 75 5 at50 5
at 25
22
Conclusion

• The complexity of modern infrastructures makes
  the modeling difficult and sometimes impossible.
• The analytical validation of algorithm is
  problematic in this context.
• Hence, it is required to use experiments to
  validate algorithms and models.
• We propose Wrekavoc a tool for emulating
  heterogeneity.
• Controllable configurations of the nodes of an
  homogeneous cluster
• Independent degradation of the characteristics
• Allows for a quantitative comparison of
  algorithms.