GangSim: A Simulator for Grid Scheduling Studies

1
GangSim A Simulator for Grid Scheduling Studies

• Catalin L. Dumitrescu
• The University of Chicago

Ian Foster Argonne National Laboratory The
University of Chicago
2
Talk Outline / Part I

• Part I
• Introduction
• Our Approach GangSim, a discrete simulator
• Motivating Scenarios
• Architecture
• Evaluation Criteria
• Part II
• Simulation and Validation Results
• Conclusions and Questions

3
Introduction

• Large distributed Grid systems pose new
  challenges
• Overwhelming resource characteristics
• Complex workload characteristics
• Complex interactions and resource allocations
• Analytical modeling is either impractical or
  impossible

4
Our Approach GangSim

• Derived from Ganglia Monitoring Toolkit
• Real-time simulator
• Focus on local VO interactions
• Mixing simulations with real testbeds
• Provide simple means for result visualization
• Interactions with various Resource Managers (RMs)
  

5
GangSim Novelty

• Simulates (and Handles)
• Sites with RMs
• VO and groups
• Submission hosts
• Model usage allocations (SLAs) at several levels
• Capacity to combine simulated results with real
  results collected from a real Grid
• Useful for simulations of future trends

6
Environment Overview
7
Environment Details

• Simulations target environments with
• large number of resources
• resource owners
• VOs
• A few examples are
• Grid3
• OSG
• TeraGrid
• DataGrid

8
Initial Research Problems

• What site usage policies are appropriate in a
  Grid environment, and how do these policies
  impact achieved site and VO performance?
• What usage policy may be applied at the VO
  level?
• What site selection policies are best suited for
  various Grid environments?

9
GangSim Details
10
GangSim Concepts

• Site characterized by various metrics about CPU,
  disk space and network connectivity
• VO composed of a groups and users
• External Schedulers, Local Schedulers, and Data
  Schedulers scheduling decision points at various
  levels in the grid
• Policy enforcement points (S-PEP and V-PEP)
  responsible to gather usage and allocation
  information and provide/control how many jobs
  should run

11
GangSim Strategies

• Various algorithms can be used for scheduling
• Site usage policy
• Simple fair share
• Extensible fair share
• Commitment fair share
• Others
• ES task assignment strategies
• Last recently used (according to available
  allocations)
• Least used (according to available allocations)
• Round robin / random assignment ()

12
Implementation Details

• Ganglia (and VO-Centric Ganglia) various
  components were replaced
• New components
• Simulator modules track client and provider
  states
• Task assignment policies various algorithm
  invoked during running
• Metric aggregators monitoring sub-components
  used for scheduling decisions
• Grid components internal data structures
• Interfaces a set of CGI scripts remotely
  accessible

13
Interface Screenshot Example
14
Talk Outline / Part II

• Part I
• Introduction
• Our Approach GangSim, a discrete simulator
• Motivating Scenarios
• Architecture
• Evaluation Criteria
• Part II
• Simulation and Validation Results
• Conclusions and Questions

15
Achievable Results

• Interested in three main aspects
• Task Assignment and Policies
• Simulated Architecture Variations
• Simulator Performance

16
Task Assignment and Policies
Round Robin Assignment Policy
Least Used Site Assignment Policy
Round Robin Assignment Policy
Used Site Assignment Policy
17
Analytical Results

• Automated performance metric computation
• Example
• ART Si1..N RTi / N

Table 2 Unsynchronized Workloads ART
Table 1 Synchronized Workloads ART
Policy/Limit No limit Fix-limit Ext-limit
Round Robin 11.09 19.39 11.32
Least Used 13.25 15.14 15.06
Policy/Limit No limit Fix-limit Ext-limit
Round Robin 7.78 14.82 9.34
Least Used 10.57 13.68 11.37
18
Simulated Architectures

• Various architectures can be simulated
• Required changes of a few parameters
• New algorithms can be considered

Analytical Approach in Site Selection
Observational Approach in Selection
19
Simulator Performance

• Important to find simulator limits
• 15 VO and 100 sites on a single GangSim instance
  is achievable

15 VOs and 100 sites (6 VOs drawn)
20
Validation Results

• Results Comparison GangSim vs. Grid3
• Site Level Comparisons
• VO Level Comparisons
• Quantitative Comparisons

21
Site Level Comparisons

• GangSim and Grid3 on a single site (FermiLab)
• 4 identical workloads
• The GangSim and FermiLab executions both
  completed in close to the same time, but show
  rather different execution behavior

Per-VO, FermiLab (Grid3)
Per-VO, FermiLab (GangSim)
22
VO Level Comparisons

• GangSim and Grid3 runs across 12 sites
• Starting times iVDGL-1 at 20 seconds, BTEV-1 and
  USATLAS-1 at 200, LIGO-1 at 700 sec, BTEV-2 at
  800, iVDGL-2 at 1000, USATLAS-2 at 1500, and
  LIGO-2 at 1700.

Per-VO, 12 sites (Grid3)
-VO, 12 sites (GangSim)
23
Quantitative Comparisons

• aggregated resource utilization (ARU)
• average response time (ART)
• ART Si1..N RTi / N.
• average starvation factor (ASF)
• ASF S ( MIN (STi, RTi) ) / S (ETi)

Table 3 Simulation (S) vs. Grid3 (G) Metrics
Level Site Site VO VO
Metric S G S G
ARU 0.12 0.16 0.07 0.06
ASF 2.36 3.9 9.88 5.09
ART 1521.31 1100.7 1824.25 639.5
24
Conclusions about GangSim

• a Grid simulator for analysis of different
  scheduling policies in a multi-site and multi-VO
  environment
• Designed for discrete simulation techniques and
  modeling of important system components
• demonstrated by describing studies of different
  VO-level scheduling policies in the presence of
  different local site resource allocation policies

25
Addressed Questions

• What site usage policies are appropriate in a
  Grid environment, and how do these policies
  impact achieved site and VO performance?
• What usage policy may be applied at the VO
  level?
• What site selection policies are best suited for
  various Grid environments?

26
Thanks

• Questions?