[Prev][Next][Index]

Second Year Project Presentation



                           DUKE UNIVERSITY
                     COMPUTER SCIENCE DEPARTMENT
                 Ph.D. SECOND YEAR PROJECT PRESENTATION

		       Trapeze: Safe At Any Speed
   
                          Kennneth Grant Yocum


                        Advisor: Dr. Jeff Chase


                             LSRC, D344
                        Tuesday, May 11, 1999
                              1:00 p.m.

Abstract:

Gigabit networks and current PCI bus technology promise data transfer rates
an order of magnitude faster from a remote node than from a local disk.
Trapeze defines an interface to Myrinet, a gigabit network, and is used as
the communication substrate for GMS, the Global Memory Service, Slice, our
new network store, MPI and CVM ports, and high-performance zero-copy
TCP/IP.  Primarily designed for kernel-to-kernel block I/O traffic, Trapeze
supports cluster computing environments by delivering bandwidth in excess
of 112 MB/s and latency less than 113 $\mu$s for 8192 byte packets.  At
this level of network performance, any overhead in the host CPU or NIC can
adversely effect end-to-end application run times.

In order to maximize application performance, the network must provide low
latency for faults of virtual memory pages, file blocks, or IP packets with
large MTU's, but also deliver high bandwidth for packet streams.
Applications must also have enough CPU cycles both to service the network
and to process data at gigabit speeds; the reduction of overheads in the
network layer is paramount to providing high-performance cluster services.

This talk explores techniques Trapeze incorporates to reduce overheads and
provide the communication power that delivers the promise of cluster
computing.  Adaptive message pipelining, AMP, balances the competing
demands of latency and bandwidth for medium to large packets. Host
overheads are reduced by incorporating scatter/gather facilities, cacheable
control register and DMA control message receives, dynamic memory pinning,
and interrupt suppression.  The talk will explore the implementation and
performance implications of these advances in relation to our work in
network memory (GMS), network storage (Slice), and zero-copy TCP/IP.
-- 
Department of Computer Science, Duke University, Durham, NC 27708-0129
Internet:	USER@cs.duke.edu