IEEE Transactions on Computers - Special issue on cache memory and related problems
Scalable Shared-Memory Multiprocessing
Scalable Shared-Memory Multiprocessing
Adaptive Routing for Dynamic Applications in Massively Parallel Architectures
IEEE Parallel & Distributed Technology: Systems & Technology
The AzusA 16-Way Itanium Server
IEEE Micro
Hi-index | 0.00 |
System partitioning provides the users of high-performance parallel servers with the flexibility in resource allocation and dynamic reconfiguration as well as fault isolation. However, the bandwidth of links that connect different domains can be wasted while links within the same domains are congested. In this paper, we present a routing mechanism that can utilize the bandwidth of otherwise unused links to balance the message traffic and lead to lower message latencies for the latency-sensitive transactions. The performance of the proposed routing mechanism was studied using an analytical model with on-line transaction processing type workload parameters. The results indicated the proposed routing mechanism reduced the congestion on the direct paths significantly and lowered the queuing delay for the links. For example, when a 4-cluster system with a fully connected network with the bandwidth of 3.2GB/s per link is partitioned into two 2-cluster domains, the queuing delay was reduced from 53ns to 37ns and resulted in the improvement of CPI by 2%.