by Sheau-ru Tong, David H. C. Du
ftp://ftp.cs.umn.edu/dept/users/du/papers/interconnect.ps
Add To MetaCart
Abstract:
Multi-stage Interconnection Networks (MINs) have been demonstrated to be one of the most cost-effective and useful communication media between processors and memory modules in parallel processing systems. In the last decade, several multiprocessor systems have been built based on this architecture. Though MINs are fairly flexible in handling varieties of traffic loads, the performance will be degraded by hotspot traffic. This is commonly known as tree saturation effect. This situation becomes worse with increasing system size. The existing combining network approach, by NYU Ultracomputer, is considered to be one of the most effective ways to alleviate the tree saturation problem. However, the combining network approach offers no performance advantage for non-hotspot traffic. This paper proposes a new approach to alleviate the tree saturation problem and at the same time improve performance of the MIN. In the proposed approach, each switch has certain packet diverting capability. Packet diverting allows packets from an intermediate switch element to bypass several subsequent stages and join with regular paths at a later stage. This helps to reduce the inner switch blocking in the MIN. This diverting capability is coupled together with a limited combining capability. Limited combining means hotspot packets are combined by the combining logics deployed at certain stages. The proposed scheme is shown to outperform the existing combining network approach.
Citations
|
186
|
Hot-spot Contention and Combining in Multistage Interconnection Networks
– Pfister, Norton
- 1985
|
|
92
|
Distributing Hot-Spot Addressing in LargeScale Multiprocessors
– Yew, Tzeng, et al.
- 1987
|
|
72
|
On a class of multistage interconnection networks
– Wu, Feng
- 1980
|
|
63
|
The IBM research parallel processor prototype (RP3): Introduction and architecture
– PFISTER, BRANTLEY, et al.
- 1985
|
|
34
|
The Monarch parallel processor hardware design
– Rettberg, Crowther, et al.
- 1990
|
|
19
|
Synchronized and Asynchronous Parallel Algorithms for Multiprocessors
– Kung
- 1976
|
|
16
|
The onset of hot spot contention
– Kumar, Pfister
- 1986
|
|
9
|
A Novel Strategy for Controlling Hot Spot Contention
– Ho
- 1989
|
|
9
|
The effectiveness of combining in shared memory parallel computers in the presence of `hot spots
– Lee, Kruskal, et al.
- 1986
|
|
7
|
A VLSI Combining Network for the NYU Ultracomputer
– Dickey, Kenner, et al.
- 1985
|
|
6
|
Preventing Congestion in Multistage Networks in the Presence of Hotspots
– Dias, Kumar
- 1989
|
|
2
|
et al.,"The NYU Ultracomputer-Designing a MIND, Shared Memory Parallel Machine
– Gottlieb
- 1983
|
|
2
|
A Cost-Effective Combining Structure for LargeScale Shared-Memory Multiprocessors
– Tzeng
- 1992
|
|
1
|
A Conceptual Approach to General Purpose Parallel Computer Architecture, "Ph.D
– Cohn
|
|
1
|
and J.R.Jump, "Analysis and Simulation of buffered Delta Networks
– Dias
- 1981
|
|
1
|
D.Klappholtz, " A Large Scale Homogeneous, Fully Distributed Parallel Machine
– Sullivan, Bashkow
- 1977
|
|
1
|
Alleviating the Impact of Tree Saturation on Multistage Interconnection Network Performance
– Tzeng
- 1991
|
|
1
|
Improving Multistage Network performance under Uniform and Hot-spot Traffics
– Peir, Lee
- 1990
|
|
1
|
A Combining Omega Network: Performance vs Implementation
– Wong
- 1986
|
