University of Oxford Parallel Applications Centre

The Oxford BSP Toolset and Profiling system (version 0.6beta)

16th June 1997: The first official release of the library is available. See the web page here.

We are pleased to announce a new release of the Oxford BSP Toolset which conforms to the BSPWW BSPlib proposal (programs written for the Oxford BSP library version 1.2 will also work with the toolset). Sources and binaries are freely available by anonymous FTP; details below.

We run mailing lists for toolset users; to subscribe, send mail to bsplib-users-request@comlab.ox.ac.uk. Bug reports should be sent to bsplib-bugs@comlab.ox.ac.uk.

Table of contents

1. How to get the toolset
2. Available platforms
3. What's new in this release
4. Coming soon
5. Schedule for next release

How to get the toolset

This release is available by anonymous FTP in the directory ftp://ftp.comlab.ox.ac.uk/pub/Packages/BSP/:

Available platforms

• Silicon Graphics Power Challenge (Irix 6.x) [SGICHALL64]
• Silicon Graphics workstation (Irix 5.x) [SGI]
• IBM SP2 running AIX [SP2]
• Cray T3D [CRAYT3D]
• Parsytec Explorer [PARSYTEC]
• Convex SPP [CONVEXSPP]
• SunOS 4.1.x [SunOS]
• SunOS 5.x [Solaris]
• PC based Unix [LINUX]
• Digital's alpha boxes [OSF1]
• Hitachi SR2001 [HITACHIMPP]

• System V shared memory [SHMEM_SYSV]
• SGI specific shared memory operations [SHMEM_SGI]
• Message Passing Interface [MPASS_MPI]
• IBM's message passing system [MPASS_MPL]
• Message passing on the Parsytec GC [MPASS_PARIX]
• Express (Parasoft Ltd) message passing [MPASS_EXPRESS]
• Cray Shmem one-sided communications [DRMA_SHMEM]

What's new in this release (0.6beta)

1. Inquiry functions for s,l,g, and nhalf. Remember to #include "bsp_level1".
2. New collective communications (bsp_bcast, bsp_fold, bsp_scan) that adapt depending upon the value of a machines l and g. The prototypes are:
   • void bsp_bcast(int frompid,void *src, void *dst, int nbytes); where
     1. src is the data to be broadcast on frompid
     2. dst is the destination of the broadcast data on all processors. dst has to be registered
     3. nbytes is the size of the broadcast data
   • void bsp_fold(void (*op)(void*,void*,void*,int*),void *src, void *dst, int nbytes);
     where
     1. bsp_fold((+),&x,&y,sizeof(int)) produces y_i = x_0 + ... + x_{p-1} on process i (where 0op is an associative operator with prototype: op(void *res, void *left, void* right, int *size); that takes two values left and right of size, and produces a result in the data-structure res. When designing op, there is no need to consider the situation where res, left, or right are aliased.
     2. src is the data-structure local to each process to be folded.
     3. dst is the destination of the fold on all processors. dst has to be registered.
     4. nbytes is the size of src. It has to be the same on all processors.
   • void bsp_scan(void (*op)(void*,void*,void*,int*), void *src, void *dst, int nbytes);
     where bsp_scan((+),&x,&y,sizeof(int)) produces y_i = x_0 + ... + x_i on process i (where 0<i< p)
     bsp_scan is therefore similar to fold, except a parallel prefix is performed in preference to reduce.
3. An implementation for the T3D. It hasn't been fully tested yet, so it probably contains a few bugs :-(
4. Extra command line arguments to bspfront/bspcc/bspf77/bspc++. In particular, the toggles for sanity checking and resource friendly barriers have been separated from the -O{1,2,3} optimisations:
   -flibrary-level <integer> There are three different levels of library available:
   • 0 Sanity checking version of the library that performs checks on registration, out of bound DRMA, and consistent arguments to collective communications. The communication primitives at this level of optimisation may not be optimal, but are guaranteed to be resource friendly (see -flibrary-level 1 below).
   • 1 Resource friendly version of the library. This means that if more processes (including other users processes) are running on the machine than processors, then the performance of the library doesn't deteriorate.
   • 2 At this optimisation level the library has been tuned in such a way that it assumes no other major processes are running on the parallel machine. When benchmarking, or collecting results, use this optimisation level.
5. Fixed some bugs. In particular:
   • Mark Goudreau's bug on sending a BSMP message to yourself.
   • Tim Lanfear's registration bug
   • Bruce Stephens problem when p=1 on the Cray T3D
   • Dominic Da Silva installation problem for Linux
   • SysV semaphores now work when a debugger attaches to a process.
   • Shared memory segment and semaphores are properly cleaned up in the SysV release of the library---no need for ipcrm!
6. The only known bugs at the moment are (these will be changed by next release):
   1. CRAY: Communication on the Cray has to be in word sized chunks. This will be changed into byte chunks as in the proposal.
   2. CRAY: bsp_put and bsp_get's interfere with each other, as opposed to all the gets occurring before the puts as in the proposal.
   3. GENERAL: bsp_put and bsp_get from and into the same process take effect immediately and not at the end of the superstep.

Coming soon

1. Manual pages
2. New platforms: Windows NT, Fujitsu AP1000
3. Shrinking of the profiling data output
4. Threads implementation that allows real DRMA between processes (i.e., start p threads, which will all share the same address space. Then remap their address spaces so that they are all disjoint, thus simulating standard processes. However, they will now be able to play with each others address space).
5. NOW version including ATM specific communications.

Schedule for next release

The Oxford BSP Toolset is a product of the BSP Programming Environment, a UK Engineering Physical Sciences Research Council project in the Portable Software Tools for Parallel Architectures Programme. Our partners in this enterprise are British Aerospace Sowerby Research Centre and NAG Ltd.

© Oxford Parallel 1996
Please send comments to Jon Hill Need more information? See: What's New in BSP Worldwide
Last modified October 8, 1996