Experiences: Overcoming Data Transfer Bottlenecks across SUN-Transputer Interfaces
نویسندگان
چکیده
In many computationally-intensive tasks such as medical image reconstruction for which transputers can be used, the data communication rate between a display host such as a SUN-workstation and the transputer network becomes a constraint on system performance. Our goal is to maximise the SUN-transputer data exchange rate. Vy'e use several kinds of transputers connected to the VME-bus of the host SUN and this paper explores alternatives for designing their interfaces. We also use a VME-bus memory module memory-mapped to both the SUN and the transputers. Device drivers for the transputers connected by serial links must implement the required blocking semantics of the transputer links. There are two major designs for programmed i/o serial link interfaces-"polled" interfaces, and "interrupting" interfaces which work with a process scheduler to block the invoking process until the desired condition is true. W'e measured both kinds, expecting that the transputer hard link throughput would limit the performance. However, we found that the SUN's cpu cycles limit the performance of an Research funded by the Natural Science and Engineering Research Council ofCanada, under grants #8020 and #45395, and by the Sirnon Fraser University Center for Systems Science. @ Computing Systems, Vol. 5 ' No. 2 ' Spring 1992 159 unbuffered byte interface; a SUN4/110 is only capable of transferring 457,000 bytes/second. By reducing the software overhead in the SUN using a buffered FIFO block transfer we can increase throughput to 888,000 bytes/second. Now the bottleneck is the byte-wide VME-bus access mode used for the serial links. Throughput between the SUN and the transputers, using shared 32-bit wide VME-memory, is four times the throughput of the serial links. Hence, for optimum throughput, we use the transputer links for synchronisation of data which is exchanged in bulk using the shared memory. Further performance increases are only possible using different hardware interfaces which support the VME-bus block transfer mode.
منابع مشابه
Performance evaluation of distributed transputer networks using high-performance communication links
Interface Standards (IS) [9] become available for the interconnection of distributed high performance transputer systems; for instance the Fibre Distributed Data Interface (FDDI) and the High Performance Parallel Interface (HIPPI). The FDDI with 100 Mb/s backbone speed is an exellent Interface Standard for a Transputer Link Interface (TLI) to interconnect distributed 100 Mb/s Inmos IMS T9000 li...
متن کاملDesign and Implementation of a Prolog-to-C Compiler
A WAM-based Prolog-To-C (PTC) compiler is described. It is written in the C language and used together with Paralogic's Prolog interpreter. Prolog programs may consist of several independently compiled modules. The principles of modularization are described. Mixed programming features (Prolog-C and CProlog programming interfaces) are also explained. The PTC compiler is portable and has been por...
متن کاملA Monitoring System for a Transputer-Based Multiprocessor
This paper describes a distributed performance monitoring system for a transputer-based multiprocessor, Hathi-2. The monitoring system allows the programmer to measure how well a parallel program utilizes the hardware resources of the multiprocessor system, i.e., the processor CPU’s and communication links, during program execution. It can be used for identifying performance bottlenecks in para...
متن کاملDeveloping transputer application circuits
4Gbytes physical memory external to the transputer). The transputer and the programming language Occam evolved together. Occam was designed to simplify the task of concurrent programming. An Occam program is made up of a number of processes which can be declared to run sequentially or concurrently. Concurrent processes, which cannot use shared resources, communicate across Occam channels. These...
متن کامل1 An Efficient Zero - Copy I / O Framework for UNIX ®
Traditional UNIX® I/O interfaces are based on copy semantics, where read and write calls transfer data between the kernel and user-defined buffers. Although simple, copy semantics limit the ability of the operating system to efficiently implement data transfer operations. In this paper, we present extensions on the traditional UNIX interfaces that are based on explicit buffer exchange. Instead ...
متن کاملذخیره در منابع من
با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید
عنوان ژورنال:
- Computing Systems
دوره 5 شماره
صفحات -
تاریخ انتشار 1992