OS, Message Passing, and Runtime Tools Panel

1
OS, Message Passing, and Runtime Tools Panel

• Rajeev Thakur
• Mathematics and Computer Science DivisionArgonne
  National Laboratory

2
Outline

• Burning issues in operating systems
• What is in MPI-2?
• Which of the runtime tools really work?

3
Burning Issues in OS

• What will be the impact of Linux and open source
  on large supercomputing platforms?
• Large impact
• Everyone is experimenting
• Mainstream vendors are exploring use of Linux in
  10TF systems
• Large systems in labs, e.g., Cplant and Chiba
  City
• Many medium size systems, which are growing,
  e.g., Putchongs cluster in Thailand is up to 72
  nodes
• Raised expectations for faster turnaround on
  problems

4
Burning Issues in OS

• Is Linux only for small and medium scale
  clusters?
• That is the question we all are trying to answer
• For distributed-memory systems, the network must
  scale. Linux doesnt come in the way.
• For large clusters we need cluster management
  tools, scalable scientific software, middleware
• For shared-memory systems, on the other hand, the
  OS must scale

5
Burning Issues in OS

• Will we need a version of Linux specially
  designed for large supercomputers?
• To be determined
• Module system a step in the right direction

6
What is in MPI-2?

• Extensions to the message-passing model
• Dynamic process management
• One-sided operations
• Parallel I/O
• Making MPI more robust and convenient
• C and Fortran 90 bindings
• External interfaces
• Extended collective operations
• Language interoperability
• MPI interaction with threads

7
What is in MPI-2?

• Will the one-sided operations be widely used?
• One-sided ops. are a popular programming model on
  Cray T3D and T3E
• Note that simple put/get/barrier enabled by
  hardware support on T3D/T3E
• MPI-2 one-sided ops. expose complex memory
  consistency issues (like MPI-1 exposed buffering
  issues)
• MPI-2 one-sided ops. are designed such that they
  can be implemented efficiently even where there
  is no hardware support for one-sided ops.

8
What is in MPI-2?

• one-sided ops
• expected to support libraries (e.g., Global
  Arrays)
• Example application FLASH (Univ. of Chicago)
• Original code uses PARAMESH library, which uses
  Cray shmem (put/get/barrier)
• Straighforward port to MPI-1 by using MMPI
  library, which translates Cray shmem to MPI-1
• Now have PARAMESH directly using MPI-1
• Want to investigate implementing PARAMESH with
  MPI-2 one-sided ops.

9
What is in MPI-2?

• MPI-2 one-sided ops. implementation status
• Fujitsu, NEC, Hitachi have it
• HP has released a subset
• Compaq, IBM in progress
• Coming in MPICH...

10
What is in MPI-2?

• Do new technologies such as VIA, ST, and
  Infiniband have any impact on the one-sided
  specification?
• They have impact on the implementation of both
  MPI-1 and MPI-2
• We are redesigning MPICH internals, one reason
  for which is to effectively use these new
  technologies
• Not clear, but not likely, that they have impact
  on the one-sided specification

11
What is in MPI-2?

• Will the client/server/dynamic-process operations
  change the way large production supercomputing
  machines are used?
• In the short run, no large scheduled batch jobs
  will rule for a while.
• In the long run, dynamic scheduling, distributed
  resource discovery, adaptive algorithms can use
  MPI-2 dynamic process ops.

12
Which of the Runtime Tools Really Work?

• Do we need large-scale interactive debuggers to
  work on thousands of nodes?
• Yes, at least on 100 nodes
• Some bugs, however, reveal themselves only on
  thousands of nodes
• How far does Totalview scale?

13
Which of the Runtime Tools Really Work?

• Will the open-source software movement lead to
  improved runtime tools for large supercomputers?
• Yes, it already has. Examples MPI
  implementations, PBS, PVFS, ROMIO, others
• Vendors may make tools open source

14
Which of the Runtime Tools Really Work?

• What new runtime tools can users expect to help
  improve the usability of large supercomputers?
• One example Jumpshot, a performance
  visualization tool developed at Argonne, is being
  designed to support multi-gigabyte trace files.
  (collaboration with IBM and Livermore)

15
OS, Message Passing, and Runtime Tools Panel

• Rajeev Thakur
• Mathematics and Computer Science DivisionArgonne
  National Laboratory