Time Warp on the Blue Gene

1
Time Warp on the Blue Gene

• Analysis of Time Warp on the Blue Gene
  Supercomputer.

2
Overview

• Port the ROSS kernel from shared memory to the
  Blue Gene/L
• Compare the performance of ROSS-MPI with ROSS-SMP
  on a shared memory platform
• Investigate the performance of ROSS-MPI on the
  Blue Gene/L

3
Motivation

• Distributed memory computers, scale bigger than
  shared memory computers.
• Models exist that are beyond the capability of
  shared memory computers.
• The Internet contains 500,000,000 host ISC,
  modelling the Internet on shared memory often
  requires compromises in model design.

4
Background

• Time Warp is an optimistic parallel discrete
  event simulator.
• ROSS is an efficient Time Warp implementation.
• Blue Gene/L is a distributed memory
  supercomputer, capable of 500 TFLOPS.

5
TimeWarp Architecture, Recap

• Logical Processor (LP) are tasks, the basic unit
  of work.
• Processor Elements (PE) are mapped to processors.
• Global Virtual time indicates the earliest time
  for an event within the system.
• Events older than GVT can be removed from memory.
• Events that execute out of order can be rolled
  back.

6
Approach

• Implement the Processor Elements PE as separate
  MPI tasks.
• Split or remove global data structures.
• Replace Fujimoto with an efficient distributed
  memory algorithm

7
Remote Events

• Process Elements (PE) are mapped to MPI tasks.
• Events between the Processor Elements are sent as
  MPI messages, Remote events.
• Remote events increase memory consumption.
• Remote events lead to ROSS-MPI consuming memory
  during GVT computation.

8
GVT Computation algorithm

• Delivers the consistent cut, by performing a
  global sum over the number of outstanding
  messages.
• Transient message problem avoided, as there are
  no outstanding messages when it is computed.
• Simultaneous reporting avoided, as GVT
  computation is delayed until all messages are
  received.

9
Workloads, PCS and PHOLD

• PHOLD, a synthetic benchmark that exercises the
  performance of Time Warp. Workload is due to
  event scheduling. Characterised by a random
  communication pattern.
• PCS, a model of a mobile phone network. An
  example of a typical model.

10
Performance under SMP
11
Performance under SMP

• All events were randomly scheduled at another LP.
• ROSS-MPI's performance decreased with processor
  count.

12
Performance under SMP

• Performance improvement is sub-linear with
  increasing processor count.

13
PHOLD, Scaling
14
PHOLD,Scaling

• With a population of 8 million events, peak
  performance of 300 million events/sec with 4096
  processors.
• With a population of 16 million events,
  performance of 500 million events/sec with 8192
  processors.
• PHOLD modified, only 10 percent of events are
  remote.
• Results are preliminary.

15
Conclusion

• ROSS-MPI can simulate larger problems than
  ROSS-SM.
• ROSS-SM is noticeably faster than ROSS-MPI on
  shared memory.
• The scalability and performance of ROSS-MPI on
  the Blue Gene/L is encouraging.

16
Summary

• Port the ROSS kernel to the Blue Gene/L
• Compare the performance of ROSS-MPI with ROSS-SMP
  on a shared memory platform
• Investigate the performance of ROSS-MPI on the
  Blue Gene/L