Faucets Queuing System

1
Faucets Queuing System

• Presented by,
• Sameer Kumar

2
Basic Idea

• Queuing System to manage Adaptive Jobs
• Adaptive jobs
• Jobs that can shrink and expand at runtime
• Simulations Provided encouraging results
• Also intended to be a general purpose queuing
  system that supports Generic, non-migratable
  Charm and MPI jobs

3
Adaptive Jobs

• Jobs that can dynamically increase (expand) or
  decrease (shrink) the number of processors they
  are running on
• Motivation
• Improve system utilization
• Decrease system response time
• Properties
• minpe, minimum number of processors required for
  the job, related to the memory requirements of
  the job
• maxpe, maximum number of processors, related to
  speedup
• profit, profit from running the job
• deadline, deadline before which the job should be
  finished

4
Adaptive Job Example

• Consider a 128 processor system
• Job A arrives and requests 80 Processors and is
  started on 80 processors
• Job B then arrives and requests 64 processors
• In traditional systems Job B will be queued and
  allocated 64 processors only after Job A
  finishes, while part of the system remains idle
• With adaptive jobs Job A can be shrunk to 64
  processors and Job B can be started and after job
  A finishes Job B can expand and use all the
  processors

5
Adaptive Job Scheduler

• Adaptive Job Scheduler manages adaptive jobs
• Three major components
• Job Manager
• Accepts jobs, schedules them on the parallel
  system, and frees resources when the job is done
• Scheduling Strategy
• A plug-able component that makes decisions on
  which jobs to schedule
• Database
• Logs all events that occur in the scheduler and
  can be used in case of a crash

6
Adaptive Job Scheduler
7
Performance

• Simulation Results on 64 processors with mean job
  execution time of 64.5 sec for utilization
  maximizing
• strategy
• Experiments on Linux Cluster on 64 processors and
  mean job execution time of 60 sec

?Arrival Rate MRT Mean Response
Time Utilization Processor utilization Load
Factor (lf) Execution Time?
8
Features

• Multithreaded for fast response
• Logs all job related information to a database
• This helps in crash recovery and,
• Improves security of the system
• Uses Unix sockets for communication
• Unix Sockets improve the efficiency of the system
  
• The also restrict access to the scheduler
• Provides timed termination of the jobs

9
Features (continued)

• Accepts both batch and interactive jobs
• MaxPE and MaxTime are parameters to the system
  and can be used to restrict unlimited access to
  the parallel machine
• Tested on the cool Linux cluster at PPL
• Adaptive jobs currently implemented in Charm
  and MPI
• For more details check out
• http//charm.cs.uiuc.edu/research/faucets/faucets.
  html

10
Super-user Access

• The Queuing system scheduler runs with super-user
  privileges
• When a new job arrives it is executed with the
  permissions of the user
• The code has been checked for stack overflows
• Direct Access to the parallel machine is blocked
  by removing the permissions for rsh, ssh etc
• To start a job the scheduler changes its group id
  to a Queuing System group which can access the
  cluster

11
Queuing System Commands

• Similar to current queuing systems
• fsub is the command to submit batch jobs to the
  queuing system
• frun runs jobs interactively
• fjobs lists the jobs
• fkill can be used to kill jobs

12
Conclusions and Future Work

• Queuing system has been tested and is ready to be
  installed on the Turing cluster
• Make the scheduler manage multiple heterogeneous
  clusters by supporting the concept of queues
• Some of the queues could be batch and others
  interactive
• Interactive queues would allocate multiple jobs
  to the same node depending on the utilization of
  the nodes
• Running the scheduler on SP2 and other
  multiprocessing architectures
• One of the solutions would be to run the faucets
  scheduler on top of a commercial queuing system