Copyright
Allan Gottlieb.

Comparing the NYU ultracomputer with other large-scale parallel processors online

. (page 1 of 1)
Online LibraryAllan GottliebComparing the NYU ultracomputer with other large-scale parallel processors → online text (page 1 of 1)
Font size
QR-code for this ebook


Comparing the NYU Ultracomputer With
Other Large-Scale Parallel Processors



Allan Gottlieb 1

Courant Institute, N.Y.U.

251 Mercer Street

New York, N.Y. 10012

[email protected]

floyd!cmd2!gottlieb

Ultracomputer Note #60

July, 1983

ABSTRACT

We describe the proposed NYU Ultracomputer, a shared memory MIMD
parallel machine composed of thousands of autonomous processing ele-
ments. This machine uses an enhanced message switching network having
the topology of an ft-network to approximate the ideal of conflict-free ac-
cess to a common memory and to implement efficiently a new fetch- and-
add synchronization primitive. We outline the hardware required to con-
struct such a system, consisting of 4096 processors, using 1990 technology
and refer to other work indicating how the goal of a distributed operating
system free from serial bottlenecks can be achieved by employing the
fetch-and-add primitive. Finally, we compare the Ultracomputer project
with other research in parallel processing.



: This work was supported by DOE grant DE-AO02-76ER03077 and by NSF grant NSF-
MCS79-21258.



Table of Contents

1. Introduction 1

2. Machine Model 1

2.1. Paracomputers 1

2.2. The Fetch-and-add Operation 1

2.3. The Power of Fetch-and-add 2

3. Machine Design 3

3.1. Network Design 3

3.1.1. ft-network Enhancements 3

3.1.2. Implementing Fetch-and-add 4

3.2. Local Memory at each PE 4

3.3. Machine packaging 5

4. Other Research 5

4.1. Alternate Machine Models 5

4.2. Languages 6

4.3. Granularity of Parallelism 6

4.4. Processor Count and Performance 6

4.5. Networks 7

4.6. Compilers and Local Memory 7

5. Conclusion 7

6. References 8



P¬Ђgel



shared variable and many fetch-and-add operations simultaneously address V, the
effect of these operations is exactly what it would be if they occurred in some
(unspecified) serial order, i.e. V is modified by the appropriate total increment
and each operation yields the intermediate value of V corresponding to its posi-
tion in this order. The following example illustrates the semantics of fetch-and-
add: Assume V is a shared variable, if PEi executes

ANSi


1

Online LibraryAllan GottliebComparing the NYU ultracomputer with other large-scale parallel processors → online text (page 1 of 1)