Xiaodong Zhang, Yong Yan, and R. Castaneda. Comparative performance evaluation of hot spot contention between min-based and ring-based shared-memory architectures. Technical Report 94-01-03, High Performance Computing and Software Laboratory, The UniversityofTexas at San Antonio, 1994. 12  Home/Search   Document Details and Download   Summary   Related Articles   Check

....(snoopy, linked list and full map directory) and the processor speed, and they compared a unidirectional ring architecture with a split transaction bus. Zhang et al. compared the performances of the hierarchical ring and the multistage interconnection network (MIN) under the hot spot contention [76]. They used analytical models to compare the MIN and the ring networks and validated the results with simulations on KSR 1 [34] and BBN TC2000 [5] Lang et al. studied the effective bandwidth of the crossbar switches and compared it to that of the multiple bus interconnection network using ....

....the bidirectional ring. Note that case (i) and (ii) are identical on the unidirectional ring since the unidirectional cannot exploit the nearest neighbor communication pattern of non local accesses. In case (iii) most of misses are filled by the local memory. Taking similar studies in the past [23, 36, 55, 76] into consideration, the range of the miss rate (MR) is set to be from 0.01 to 0.05. The plots for the miss latency and the processor utilization over the miss rate (miss per reference) are given in Figures 7 and 8. In case (i) since access locality is quite low, both rings exhibit high access ....

X. Zhang, Y. Yan, and R. Casta~neda, "Comparative Performance Evaluation of Hot Spot Contention Between MIN-based and Ring-Based Shared-Memory Architectures," Transactions on Parallel and Distributed Systems, IEEE, Vol. 8, No. 8, 872--886, August 1995.

....results in a given mutual algorithm exhibiting different execution patterns on each machine. ffl System reactions to hot spots. The TC2000 is much more hot spot sensitive than the KSR 1 due to different network structures. According to the experimental and analytical results reported in [12], the ring network transactions in the KSR 1 will be reduced no more than 50 in the presence of memory hot spots. This compares with a 300 latency change in the TC2000. ffl Process sequences. The rotating and slotted ring network of the KSR 1 may naturally order CS requests to be sequenced by ....

....to enter the CS in a certain pattern, such as a tree structure. Thus, the B L algorithm should have less software delay time in practice although it may have higher contention than that of the YA algorithm. However, the ring structure of the KSR 1 system may effectively handle network contention [12]. Finally, the shared variables are in the form of a vector in the B L algorithm. This structure allows us to use the cache subpage (128 bytes) which will bring up to 32 integer variables at a time to a local cache for reduction of remote accesses. Figure 8 presents process response time ....

[Article contains additional citation context not shown here]

X. Zhang, Y. Yan and R. Casta~neda, "Comparative performance evaluation of hot spot contention between MIN-based and ring-based shared-memory architectures", IEEE Transactions on Parallel and Distributed Systems, Vol. 6, No. 8, 1995.

....CC NUMA due to the overhead of more frequent data movement in the COMA system. This conclusion is consistent to the analytical results reported in Figure 4 in terms of hot spot effects. In practice, a hot spot is usually generated only by part of processors in the system. Experiments reported in [16] simulate this type of memory access patterns, which used 57 out of 64 processors in a KSR 1 system to generate the hot spot on another remote cache module, remaining 6 remote cool cache modules. The miss latencies of remote reads and remote writes of one word, one block, two blocks and three ....

....measured under an environment without any hot spots, an environment with the hot spot generated by cache references in a word unit, and an environment with the hot spot generated by cache references in a block unit. In comparison between the fixed and movable hot spot experiments, the results in [16] indicate that a movable hot spot slightly increases the access delay to 23 hot spot rate 0.1 0.2 0.3 0.4 0.5 0.6 w miss in COMA 30.2 33.3 36.1 39.2 42.1 45.2 w miss in NUMA 34.4 33.8 33.3 33.6 34.1 33.7 r miss in COMA 25 25.1 25.2 25 25.1 25.2 r miss in NUMA 25.3 25.2 25.5 25.1 25.2 25.5 ....

X. Zhang, Y. Yan and R. Casta~neda, "Comparative performance evaluation of hot spot contention between MIN-based and ring-based shared-memory architectures", to appear in IEEE Transactions on Parallel and Distributed Systems.

....be examined precisely by program execution on a particular architecture. By tracing program executions on the Exemplar, we present overall scaling capability of the architecture. We also compared the system and application performance with the performance of the KSR 1 reported in [13] 14] and [15]. The testing programs on the Exemplar were run within the CXpa (CONVEX Performance Analyzer) 2] environment for collecting the execution timing results. The CXpa is a software performance monitor which has the capability to time loops and count cache misses. Although the profiler intruded ....

....degrade all network traffic, not just the traffic to the shared variables. The effect is defined as tree saturation, where traffic to the hot memories backs up at the switch and interferes with other traffic, including that to non hot memories. Furthermore the modeling and experimental work in [15] indicates that blocking MINs are much more hot spot sensitive than slotted rings although the blocking MINs may provide a fast connection for memory access. Therefore, hot spot tests are important to show how tolerable a network structure to non uniform network traffic patterns. Comparative ....

[Article contains additional citation context not shown here]

X. Zhang, Y. Yan and R. Casta~neda, "Comparative performance evaluation of hot spot contention between MIN-based and ring-based shared-memory architectures", IEEE Transactions on Parallel and Distributed Systems, Vol. 6, No. 8, 1995, pp. 872-886.

No context found.

Xiaodong Zhang, Yong Yan, and R. Castaneda. Comparative performance evaluation of hot spot contention between min-based and ring-based shared-memory architectures. Technical Report 94-01-03, High Performance Computing and Software Laboratory, The UniversityofTexas at San Antonio, 1994. 12

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC