S. Scott and G. Sohi. The Use of Feedback in Multiprocessors and its Application to Tree Saturation Control. IEEE Transactions on Parallel and Distributed Systems, 1(4):385--398, October 1990.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....due to an effect they call tree saturation. They also found the technique of message combining to be an effective means of eliminating this problem if it arises due to lock or synchronization contention. Combining and feedbark schemes have been suggested as partial solutions to the problem [18 20]. Atiquzzaman [10] proposed an efficien: Markov chain modeI for the performance evaluation of a single buffered Omega switch in the presence of a hot spot Three different buffer types for 2 x 2 SEs are ana lyzed in [21] for the unbounded queue size and queue s ze equal to one. The aim of this ....

S.L. Scott & G.S. Sohi, The use of feedback in multiprocessors and its applications to tree saturation control, IEEE Trans. on Parallel and Distributed Systems, 1990, 943-948.

....but none of which we are aware consider interconnection network utilization. Another class of adaptive protocols includes the COMA [9, 13] and R NUMA [8] protocols. These protocols migrate data to where it is used, adaptively reducing communication and network traffic. Networks. Scott and Sohi [27] proposed using network feedback to adaptively avoid tree saturation in multistage interconnection networks. They use feedback control theory to adjust network usage to avoid performance degradation due to hot spots in memory access patterns. This work differs from bandwidth adaptive snooping in ....

S. Scott and G. Sohi. The Use of Feedback in Multiprocessors and its Application to Tree Saturation Control. IEEE Transactions on Parallel and Distributed Systems, 1(4):385-- 398, Oct. 1990.

....under a variety of realistic workloads has been studied for a long time and this resulted in improved switch and interconnect designs. One of the problems that has been the hardest to solve is that of congestion management in the face of unpredictable traffic patterns. Feedback techniques [7] such as multi lane backpressure [5] have been experimented with and the results are promising. The flow control techniques in a GigaNet based interconnect are novel in that (1) the feedback is directly to the VCI source and not to the predecessor switch in the path, 2) it does not employ any ....

Scott, S.L.; Sohi, G.S., The use of feedback in multiprocessors and its application to tree saturation control, IEEE Transactions on Parallel and Distributed Systems, Volume: 1 4, Oct. 1990 , Page(s): 385 --398.

....the performance of such networks. Performance of crossbar multiprocessors, unbuered MINs and buered MINs in the presence of hot spots have been developed by Atiquzzaman [2326] Combining and feedback schemes have been suggested as solutions to the problem in multistage interconnection networks [22, 27, 28]. The bandwidth of a crossbar system under hot spot trac conditions have been determined in Ref. 23] The model assumes processors with uniform priority. Analytical modeling permits one to obtain a quick evaluation of the expected performance of the system as a function of the dierent ....

Scott SL, Sohi GS. The use of feedback in multiprocessors and its applications to tree saturation control. IEEE Trans Parallel Distrib Syst 1990;1(4):38598.

....communication between the forward and return paths of the network, as does combining. Feedback from the memory modules can be used to block messages destined for a hot memory module from entering the network, whenever the length of the queue at the memory module is above a certain threshold. In [123], Scott and Sohi showed that this method can improve bandwidth and decrease latency for messages destined to cold memory modules, as long as only a fraction of the processors are hot , that is, are accessing the hot memory module at a rate greater than that of uniform traffic. In their ....

....unbounded buffers, to the latency of Type A and Type B switches. Figure 2.12 shows a design similar to the Type C switch, but with the two single input queues preceded by instead of followed by a crossbar. Such a design, often with a buffer at the input, has frequently been proposed (see, e.g. [29, 123, 144]. If data movement from chip to chip takes only a single cycle, this design shows the same behavior as Type C switches, as can be seen by redrawing the network so that the queues are part of the same switch as the crossbar (see the dotted box in Figure 2.12. The buffer at the input increases the ....

Steven L. Scott and Gurindar S. Sohi. The use of feedback in multiprocessors and its application to tree saturation control. IEEE Transactions on Parallel and Distributed Computing, 1(4):385--398, 1990.

....requires tuning the appropriate time outs, and when the timeouts are tuned for robustness at higher loads, there is a performance penalty for light loads. Scott and Sohi describe the use of explicit feedback to inform nodes when tree saturation is imminent in multistage interconnection networks [24]. This approach also requires tuning of thresholds. The technique proposed by Kim et al. 16] allows the sender to kill any packet that has experienced more delays than a threshold. This approach pads shorter packets to ensure that the sender can kill a packet at any time before its first flit ....

S. Scott and G. Sohi. The Use of Feedback in Multiprocessors and its Application to Tree Saturation Control. IEEE Transactions on Parallel and Distributed Systems, 1(4):385-- 398, October 1990.

....The network is overloaded and even data not destined for the hot spot is delayed substantially. Several concepts to alleviate saturation tree effects on network performance have been proposed. These concepts can be divided into three main classes: combining techniques [6] flow control techniques [7], and enhanced switch box designs [3, 10] In combining techniques, several messages destined for the same destination are combined into a single message to reduce the number of hot messages within the network. This technique results in high hardware overhead, and might fail if hot spot messages ....

S. L. Scott and G. S. Sohi, "The use of feedback in multiprocessor and its application to tree saturation control," IEEE Trans. Par. and Distr. Syst., Vol. 1, No. 4, October 1990, pp. 385-398.

....interconnection network. Early work in the performance analysis of unbuffered MINs was done by Patel [1] Performance of unbuffered MINs in the presence of nonuniform traffic have been reported in [2, 3] Performance of MINs using buffered SEs under uniform and non uniform traffic are presented in [4, 5, 6, 7, 8, 3]. Three different 2 Theta 2 crossbar switching elements (SE) are analyzed in [9] for the single and unbounded queue sizes. The aim of this paper is to study the performance of MINs having finite buffered SEs, in the presence of uniform and hot spot traffic patterns. The results of the research ....

S.L. Scott and G.S. Sohi, "The use of feedback in multiprocessors and its applications to tree saturation control," IEEE Trans. on Parallel and Distributed Systems, pp. 943--948, Oct. 1990.

....h (N 1) 4) The request rate of hot processors r 1 is, however, limited to r 1 = 1 (1 f) r 2 f (1 h (N 1) 1 1 h (N 1) 8 When there is tree saturation, the system behaves as if all, not just fN, processors are participating in the hot spot activity. It has been conjectured in [Scot90] that the system behaviour under tree saturation can be modelled as if all N processors are generating hot spot requests at a smaller fraction fh. Note that in reality only fN processors are generating hot spot requests at a fraction h. This conjecture has been validated by simulation experiments ....

....that the system behaviour under tree saturation can be modelled as if all N processors are generating hot spot requests at a smaller fraction fh. Note that in reality only fN processors are generating hot spot requests at a fraction h. This conjecture has been validated by simulation experiments [Scot90]. Then, the maximum bandwidth per processor b ts is obtained by substituting hf for h in Eq. 2) as b ts = 1 1 h f (N 1) 5) 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Fraction of hot processors, f Maximum bandwidth per processor (requests cycle) bts bnts 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 ....

[Article contains additional citation context not shown here]

S. L. Scott and S. S. Sohi, "The Use of Feedback in Multiprocessors and Its Application to Tree Saturation Control ," IEEE Trans. Parallel and Distributed Systems, Viol. 1, No. 4, October 1990, pp. 385-398.

....and having leaves at the processors, obstructs the flow of hot traffic, as well as any cold traffic that has to cross the saturated tree. Tree saturation results in degradation in the performance of the network. Combining and feedback schemes have been suggested as partial solutions to the problem [20, 21, 22]. The degradation of an unbuffered network in the presence of a hot spot is due to the high rate of contention at SEs carrying the hot traffic. Performance of MINs in the presence of hot spot traffic is usually obtained from simulation studies. Unfortunately, such simulation studies are very ....

S.L. Scott and G.S. Sohi, "The use of feedback in multiprocessors and its applications to tree saturation control," IEEE Transactions on Parallel and Distributed Systems, vol. 1, no. 4, pp. 385--398, October 1990.

....case of uniform memory reference pattern has been reported in [2, 3] Favorite memory case has been analyzed in [4] However, the assumption of uniform memory reference pattern in a multiprocessor system is not realistic. Different non uniform traffic patterns have been reported in the literature [5, 6, 7, 8]. Hot spot traffic is a particular type of non uniform traffic which arises due to the use of synchronization primitives and shared variables in a shared memory system. The effect of hot spots have been found to degrade the performance of multiprocessor systems using crossbar [9] multiple bus ....

S.L. Scott and G.S. Sohi, "The use of feedback in multiprocessors and its applications to tree saturation control," IEEE Transactions on Parallel and Distributed Systems, vol. 1, no. 4, pp. 385--398, October 1990.

....included in the initial design, is shown to bring the system quite near to this ideal. Traffic throttling is not a new idea. In the WAN world it has long been commonplace for network routers to examine round trip latencies and throttle back traffic on congested routes accordingly. Scott and Sohi [4] as well as Farrens, Wetmore and Woodruff [5] propose using feedback from hot memory modules to restrict the injection of traffic destined for those modules. We are not proposing to add any such complex mechanisms, though. Dally and Aoki [6] have suggested that since their network design is ....

S. L. Scott and G. S. Sohi, "The use of feedback in multiprocessors and its application to tree saturation control," IEEE Transactions on Parallel and Distributed Systems, vol. 1, pp. 385--399, Oct 1990.

....requests but can only service one at a time, the mean service time for these requests is increased. Requests to cold memory modules ( cold requests) can encounter full queues; therefore, the latency of the cold requests is increased, and the level of congestion in the network is heightened [PoHa89, ScSo90]. This degradation is called tree saturation and is compounded by the fact that its onset can be rapid [KuPf86] and may take a long time to alleviate. In addition, it takes very little nonuniform traffic to induce this condition [PfNo85, Ston87] # # 2x2 Switch 2x2 Switch 2x2 Switch 2x2 Switch ....

....modules. Unfortunately, because of the inherent latencies involved in the network, a new batch of requests may not arrive before the hot module services all of its queued requests (leaving the memory module idle) Scott and Sohi label this undershoot, an analogy from classic control systems theory [ScSo90]. Another problem with feedback occurs when several of the processors have saved a number of requests for a hot module. When the processors are notified that the module is again accepting requests, the stored requests will be simultaneously released by the processors, possibly flooding the ....

[Article contains additional citation context not shown here]

S. L. Scott and G. S. Sohi, "The Use of Feedback in Multiprocessors and Its Application to Tree Saturation Control", IEEE Transactions on Parallel and Distributed Systems, vol. 1:4 (October 1990), pp. 385-399.

....no cost if included in the design, is shown to bring the system quite near to this ideal. Traffic throttling is not a new idea. In the WANworld it has long been commonplace for network routers to examine round trip latencies and throttle back traffic on congested routes accordingly. Scott and Sohi [7] as well as Farrens, Wetmore and Woodruff [8] propose using feedback from hot memory modules to restrict the injection of traffic destined for those modules. We are not proposing to add any such complex mechanisms, though. More recently, Dally and Aoki [9] have suggested that since their network ....

S. L. Scott and G. S. Sohi, "The use of feedback in multiprocessors and its application to tree saturation control," IEEE Transactions on Parallel and Distributed Systems, vol. 1, pp. 385--399, Oct 1990.

....contention, the amount of tree saturation which occurs in non uniform traffic conditions is reduced, thereby lowering average request latency. 3. Simulations 3.1. The Network Model In order to test the effectiveness of this architecture, a number of modifications to the MIN simulator used in [FaWW91, ScSo90] were made and a range of simulations were performed. The topology of the network used in the CCHIME architecture is that of a data manipulator, which has been shown to be isomorphic with the flip, omega, banyan, and indirect binary n cube networks. The processors and caches are connected by a ....

S. L. Scott and G. S. Sohi, "The Use of Feedback in Multiprocessors and Its Application to Tree Saturation Control", IEEE Transactions on Parallel and Distributed Systems, vol. 1:4 (October 1990), pp. 385-399.

....accesses to a single hot spot. It is more suitable for SIMD machines. Other hardware schemes have been proposed for barrier synchronization, such as [BePo90] and [HwSh91] but they are not flexible enough to handle traffic to memory hot spots. Feedback has been proposed by Scott and Sohi [ScSo89] for avoiding tree saturation, but it does not improve the latency of hot spot accesses and is not a substitute for combining. Similarly, the intelligent allocation of hardware switch buffers (for example, Tzen91] relieves congestion, but does not address the problem of high latency hot spot ....

S. Scott and G. Sohi, "The use of feedback in multiprocessors and its application to tree saturation control," IEEE Trans. on Parallel and Distributed Systems, Vol. 1, No. 4, 1990.

....in order to slow done the balancing activities and to continue the branch bound algorithm. After some time the Delta values are decreased to the original values. To control this we use a feedback strategy. This type of control mechanism is well known in the design of engineering control systems [19]. It is a simple but extremely powerful method to prevent instability of the underlying system by keeping the outputs in a fixed range. For our purpose, it is necessary to keep the load balancing activities (e.g. the number of migrated subproblems) in a range that depends on the communication ....

S. L. Scott, G. S. Sohi, The Use of Feedback in Multiprocessors and Its Application to Tree Saturation Control, IEEE Transactions on Parallel And Distributed Systems, Vol. 1, No. 4, October 1990, pp. 385-398

....and having leaves at the processors, obstructs the flow of hot traffic, as well as any cold traffic that has to cross the saturated tree. Tree saturation results in degradation in the performance of the network. Combining and feedback schemes have been suggested as partial solutions to the problem [13, 19, 23]. The degradation of an unbuffered network in the presence of a hot spot is due to the high rate of contention at SEs carrying the hot traffic. Performance of MINs in the presence of hot spot traffic is usually obtained from simulation studies. Unfortunately, such simulation studies are very ....

S.L. Scott and G.S. Sohi. The use of feedback in multiprocessors and its applications to tree saturation control. IEEE Transactions on Parallel and Distributed Systems, 1(4):385--398, October 1990.

....hot memories, and the phenomenon as hot spot contention which was first reported by Pfister [14] In the case of a buffered network, a phenomenon called tree saturation severely degrades the performance of the network. Combining and feedback schemes have been suggested as solutions to the problem [15, 16, 17]. In the case of an unbuffered network, the performance degradation is due to high rate of contention at the switches carrying the hot memory traffic. The first objective of this paper is to evaluate the performance of an unbuffered Omega network under hot spot traffic. This will give an insight ....

S.L. Scott and G.S. Sohi, "The use of feedback in multiprocessors and its applications to tree saturation control," IEEE Transactions on Parallel and Distributed Systems, vol. 1, no. 4, pp. 385--398, October 1990.

....a particular memory location or module in a large system is known as a hot spot, and has been the focus of much research. Hardware mechanisms for combining requests in the network have been proposed [Gott83, Pfis85] as well as mechanisms to improve network performance in the face of contention [Tami88, Scot90, Lang88]. These mechanisms are not necessary under the uniform workload assumption, as the average rate of read requests to a line is independent of system size. However, certain workloads (those using barrier synchronization, for instance) may exhibit significant read contention, in which case it is ....

Scott, S. L. and G. S. Sohi, The Use of Feedback in Multiprocessors and Its Application to Tree Saturation Control, IEEE Transactions on Parallel and Distributed Systems 1(4), October 1990, 385-398. 167

No context found.

S. Scott and G. Sohi. The Use of Feedback in Multiprocessors and its Application to Tree Saturation Control. IEEE Transactions on Parallel and Distributed Systems, 1(4):385--398, October 1990.

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