J. H. Kim, Z. Liu, and A. A. Chien. Compressionless Routing: A Framework for Adaptive and Fault-Tolerant Routing. In Proceedings of the 21st International Symposium on Computer Architecture, April 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....are virtual channels, thus preventing the use of all the virtual channels for fully adaptive routing. Deadlock recovery strategies overcome this constraint but the cost associated with deadlock recovery can be high, specially when deadlocked messages are killed and injected again at a later time [26, 18]. Progressive deadlock recovery strategies, like Disha [2, 3, 4] are more efficient because they allocate a few dedicated resources to quickly deliver deadlocked packets, instead of killing them. Disha also uses dedicated resources to recover from deadlock. In this case, one central buffer per ....

J. H. Kim, Z. Liu and A. A. Chien, "Compressionless routing: A framework for adaptive and fault-tolerant routing," in Proceedings of the 21st International Symposium on Computer Architecture, April 1994.

....of this author no systematic study of the subject exists. There are some attempts at analytical modeling of wormhole switched networks [32, 39, 87] They are all characterized by imposing crude assumptions about stochastic independence. Otherwise, studies are based on simulation experiments [6, 13, 34, 35, 50, 88, 102, 103, 105, 106]. Unfortunately, the random nature of output data is often neglected in these cases. Usually little or nothing is said about the experimental design, hence it is hard to put confidence in the results. Combining the fact that analytical work rests on stochastic independence with the fact that ....

....there is a small number of flit buffers at each switch. This is indicated by the small zigzaged portions of the packets in figure 2.2. Note also that WH switching has some resemblance to circuit switching [150, 155] In fact, it is sometimes considered to be a pipelined circuit switching principle [5, 88]. The flit size is a design parameter but otherwise the flow control scheme is intimately related to the WH switching principle. In contrast, buffered packet switching (SF or VC) does not depend on any particular flow control algorithm. In fact, link level flow control can be omitted all ....

[Article contains additional citation context not shown here]

KIM, J., LIU, Z., AND CHIEN, A. Compressionless routing: a framework for adaptive and fault-tolerant routing. In Proc. of 21st Annual Symposium on Computer Architecture, Computer Architecture News (April 1994), pp. 289--300.

....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 reaches the destination. This can cause larger overheads when short messages are sent to distant ....

J. H. Kim, Z. Liu, and A. A. Chien. Compressionless Routing: A Framework for Adaptive and Fault-Tolerant Routing. In Proceedings of the 21st International Symposium on Computer Architecture, April 1994.

....of the router, and Section 5 concludes the paper with some performance figures. 2 Adaptive Recovery based Routing Deadlock recovery routing is optimistic as it permits unrestricted routing of packets in the absence of deadlock and efficiently recovers from rare cases of impending deadlock [8, 9]. Recent empirical studies confirm that deadlocks are highly improbable when routing freedom abounds [5] Due to this, the WARRP router implements the Disha Sequential [9, 10] recovery based routing algorithm which progressively redirects suspected deadlocked packets toward destinations rather ....

.... Due to this, the WARRP router implements the Disha Sequential [9, 10] recovery based routing algorithm which progressively redirects suspected deadlocked packets toward destinations rather than regressively aborting and later re introducing these packets into the network from source nodes as in [8]. Though highly improbable, multiple deadlocks can be resolved sequentially one after another. Progressive recovery from deadlocks is through a single flit sized deadlock buffer central to each router accessible from all neighboring routers. System wide, these buffers form what is collectively a ....

J. Kim, Z. Liu, and A. Chien. "Compressionless Routing: A Framework for Adaptive and Fault-tolerant Routing ". In Proceedings of the 21st International Symposium on Computer Architecture,IEEE Computer Society, pages 289-300, April 1994.

....to treat messages with temporarily inaccessible destination in another way. Such messages could be simply killed, temporarily stored in an intermediate node, destroyed and reintroduced later, routed in a different way, sent through a special virtual subnetwork, given higher priority etc. see e.g. [13, 16, 20]. Several such algorithms have already been suggested and the first experience shows their very good properties including almost constant mean latency for a large range of load and stability. ....

Kam, J. J., Liu, Z., Chien, A., "Compressionless routing: a framework for adaptive and fault-tolerant routing", IEEE Transactions on Parallel and Distributed Systems, Vol. 6, No. 5, pp. 229-244, March 1997.

....the sparsity of deadlocks in asynchronous networks. Thus, instead of avoiding deadlocks, precautions are taken to efficiently recover from them. In the compressionless approach, any established virtual channels which have not reached their destination within a given time delay are torn off [148]; this method is implemented using special test packets which follow flow control paths. Alternatively, a new better approach called Disha 1 provides alternative paths on demand to deliver deadlocked packets [11, 247] 1.6 Covered Topics and Further Reading In this study, we survey ....

Kim, J. H., Liu, Z., and Chien, A. A. Compressionless routing: a framework for adaptive and faulttolerant routing. IEEE Trans. Parallel Distrib. Syst. C-8 (3), 1997, pp. 229--244.

....in Fully Adaptive Algorithms Almost all the deadlock prevention mechanisms discussed so far use additional resources or implementations to prevent the formation of cycles and thereby avoid deadlock. However, studies have shown that potential deadlock situations are rare in multicomputer systems [6, 44, 58, 59] and it may not be cost effective to dedicate resources to handle rare events. Deadlock recovery is an alternative to deadlock avoidance or prevention. Using this concept, the messages can be routed fully adaptively, allowing the formation of cycles. A detection mechanism identifies potential ....

....that aborts a message whenever it is blocked beyond a threshold number of cycles. The message is then reintroduced into the network after a random number of cycles. It is shown that the abort and retry mechanism improves performance under a broad range of traffic conditions. Kim, Liu, and Chien [44] report an adaptive routing framework called compressionless routing (CR) that supports adaptive and fault tolerant routing for a wide variety of network topologies without using any virtual channels. A feature of wormhole routing that provides feedback in the form of flow control is exploited ....

[Article contains additional citation context not shown here]

J. H. Kim, Z. Liu and A. A. Chien, "Compressionless Routing : A Framework for Adaptive and Fault- Routing", Intl. symposium on Computer Architecture, pp. 289-300, April 1994.

....decision made in trading off routing freedom and deadlock formation. Avoidance based routing algorithms enforce certain routing restrictions in order to altogether avoid deadlocks [1, 2, 3] Recovery based routing algorithms relax routing restrictions and recover from potential deadlock situations [4, 5]. The circumstances under which either routing approach is preferable depend critically on the frequency with which deadlocks occur and the resulting effects. For instance, deadlock may be so infrequent for a particular network configuration that avoidance based routing inefficiently uses network ....

....DOR. Single cycle deadlocks in DOR (as shown in Figure 1) require circular overlap of messages. The source and destination pairs designated by some of these non uniform traffic patterns are such that this overlap is not possible. 4 Related Work Deadlock approximation schemes proposed previously [4, 5] have provided little insight into the frequency of true deadlocks. In contrast, our work presents frequencies of actual deadlock as well as their characteristics as they relate to key network parameters. CWGs and similar constructs have previously been used to statically represent connections ....

J. Kim, Z. Liu, and A. Chien. "Compressionless Routing: A Framework for Adaptive and Fault-tolerant Routing". In Proc. of the 21st International Symposium on Computer Architecture, pp 289-300, April 1994.

....from a node labeled two to a node labeled one. The maximum number of such moves on a k ary n cube is dnbk=2c=2e, so dnbk=2c=2e 1 virtual channels per physical channel are required. Rather than providing additional virtual channels to support deadlock free routing on a torus, Kim, Liu, and Chien [55] use an abort retry mechanism to terminate messages in a deadlock configuration. Since wormhole routing is used, messages are 51 padded with extraneous data that follows the header through the network. Enough padding is provided so that the message header reaches the destination before all the ....

J. H. Kim, Z. Liu, and A. A. Chien. Compressionless Routing: A Framework for Adaptive and Fault-tolerant Routing. In 21 st Annual International Symposium on Computer Architecture, pages 289--300, 1994.

.... routing on the offset cube topology (Section IV A) and develops the necessary routing constraints for practical, deadlock free routing protocols (Sections IV B IV D) using techniques presented in [15, 19] While recently proposed detect and recover schemes such as Compressionless Routing [22] and DISHA [3] can be readily adapted to the offset cube, they are not considered here. A. Adaptive Routing Fundamentals for the Offset Cube Because the offset cube is a multipath topology, adaptive routing can be exploited to improve both throughput and fault tolerance. At each node visited by ....

J. H. Kim, Z. Liu, and A. Chien, "Compressionless Routing: A Framework for Adaptive and Fault-tolerant Routing," Proceedings of the 21st Annual International Symposium on Computer Architecture, pp. 289-300, 1994.

....This together with the fact that each normal hop takes a message closer to the destination proves that messages are correctly delivered and livelocks do not occur. ut If faults occur during network operation, deadlocks can be avoided by draining the messages on faulty links using kill signals [13] and by removing messages destined to faulty nodes. A message, say, M , destined to a faulty node will eventually become a column message, say, an NS message, with our misrouting logic. Upon further routing, M will reach a point where it has just completed misrouting by reaching a south row of the ....

J. H. Kim, Z. Liu, and A. A. Chien, "Compressionless routing: A framework for adaptive and fault-tolerant routing," in Proc. 21st Ann. Int. Symp. on Comput. Arch., pp. 289--300, 1994.

....results were presented in part at the 24th Annual International Conference on Parallel Processing, August 1995. 2 the network to be able to dynamically route messages along alternative, possibly non minimal paths. Techniques for adaptive routing have been proposed as means of avoiding faults [4,5,9,14,19,21,24]. Fault tolerant routing algorithms have been developed for both wormhole switched and packet switched networks. These techniques generally require substantial hardware support within the network routers, but can deliver robust inter processor communication performance in the presence of a ....

....a large number of faults. However, the occurrence of three faulty components within a period of tens of hours corresponds to quite a high fault rate. Therefore the results are felt to be of practical value. For much higher fault rates or larger mean times to repair, more robust techniques exist [19,24,4,12] that make use of improved routers and associated routing techniques. Finally, the proposed techniques do accommodate a range of fault patterns: more than previously proposed techniques for fault tolerant wormhole switching. For example, many combinations of concave fault regions are accommodated ....

[Article contains additional citation context not shown here]

J. Kim, Z. Liu, and A. A. Chien, "Compressionless routing: a framework for adaptive and fault tolerant routing," Proceedings of the 21st International Symposium on Computer Architecture, pp. 289300, April 1994.

....techniques. Much of this complexity results from resolving the problem of deadlocks in the network. Routing algorithms can be made deadlock free in one of two ways. The first is by allowing deadlocks to occur and then to recover. Recovery schemes are used in both the compressionless router [KLC94] which uses an abort and retry technique similar to double buffering [W 88] and in the Disha router [KP95] which has a special set of buffers which a deadlocked message may access exclusively to reach its destination. The alternative is deadlock avoidance, which is a more commonly used ....

J.H. Kim, Z. Liu, and A.A. Chien. Compressionless routing: a framework for adaptive and fault-tolerant routing. In Proc. of the Intl. Sym. on Computer Arch., pages 289--300, 1994.

....becomes available. Therefore, the total message delivery time from source to destination depends not only on the network bandwidth but also on network traffic conditions. In many cases, when a message header arrives at the destination node, part of the message remains in memory at the source node [17]. The same is true of a message that is blocked in the network. If such a situation occurs in a one port architecture, then the source node is prevented from sending any other message until the blocked message releases the output port. In a multi port architecture, once the second message has been ....

J. H. Kim, Z. Liu, and A. A. Chien, "Compressionless Routing: A Framework for Adaptive and Fault-tolerant Routing," in Proceedings of the 21st International Symposium on Computer Architecture, pp. 289--300, 1994.

....of the k fault tolerance algorithm. It is possible to conceive implementation of the proposed algorithm using various wormhole routing protocols. The discussion in this report, however, utilizes a variant of the the fault tolerant compressionless routing (FCR) recently proposed by Kim et al. [8]. The next section briefly describes FCR. The subsequent sections describe the routing protocol used by our algorithm, and the proposed implementation of the k fault tolerance algorithm. 11.1 Fault Tolerant Compressionless Routing [8] FCR tolerates failures in routing, i.e. a message is ....

....compressionless routing (FCR) recently proposed by Kim et al. 8] The next section briefly describes FCR. The subsequent sections describe the routing protocol used by our algorithm, and the proposed implementation of the k fault tolerance algorithm. 11.1 Fault Tolerant Compressionless Routing [8] FCR tolerates failures in routing, i.e. a message is delivered to its destination in spite of few link (or router) failures. FCR is not designed to recover from processor failures, however. FCR is used to send messages between the nodes. The basic idea behind FCR is to pad, if necessary, a ....

J. H. Kim, Z. Liu, and A. A. Chien, "Compressionless routing: A framework for adaptive and fault-tolerant routing," in Int. Symp. Comp. Arch., pp. 289--300, 1994.

....or on recovering from deadlock. Avoidance based routing algorithms enforce certain routing restrictions in order to altogether avoid deadlocks [1, 2, 3] Recovery based routing algorithms relax routing restrictions but must provide a means for recovering from potential deadlocks that may form [4, 5, 6]. Thus, the main dis This research was supported in part by an NSF Research Initiation Award, grant ECS 9411587, and an NSF Career Award, grant ECS9624251. ....

....that deadlocks are possible. We wish to identify which approach is more effective at reducing cyclic non deadlocks and concomitant correlated message blockingassuming sufficient routing freedom is provided to make deadlock highly improbable (as assumed by deadlock recovery based routing schemes [4, 5, 6]) We compare Duato s avoidance based adaptive routing algorithm [3] with Disha [5] a recently proposed recoverybased adaptive routing algorithm. Three and four virtual channels are assumed for both, along with the other default network parameters. Duato s algorithm restricts routing on two of ....

[Article contains additional citation context not shown here]

J. Kim, Z. Liu, and A. Chien. "Compressionless Routing: A Framework for Adaptive and Fault-tolerant Routing". In Proc. of the 21st International Symposium on Computer Architecture, pp 289-300, April 1994.

....deadlock freedom in several ways. The first is by allowing deadlocks to occur and then to recover. This method is motivated by the observation that deadlock is a rare event, and hence should not require significant routing resources. Recovery schemes are used in both the Compressionless router [KLC94] and in the Disha router [KP95] The former uses an abort and retry technique, similar to double buffering [W 88] in which a message tries to acquire all the buffers between its source and its destination. If the message is successful, it will be delivered in time proportional to the length ....

....padded. If after some time, the message does not reach its destination, the source sends a kill signal along the message path causing the delivery attempt to be aborted. The aborted message is dropped 13 from the network, and another try is made at a later time. The Compressionless algorithm [KLC94] is adaptive, fault tolerant, and does not require virtual channels, though it is more complex than traditional routers. If message lengths are short, bandwidth may be wasted on message padding. Compressionless routing, however, is one of the few adaptive routers that provides in order delivery. ....

[Article contains additional citation context not shown here]

J.H. Kim, Z. Liu, and A.A. Chien. Compressionless routing: a framework for adaptive and fault-tolerant routing. In Proc. of the Intl. Sym. on Computer Arch., pages 289--300, 1994.

....not properly resolved, results in the permanent blocking of messages being routed. Previous studies have not provided insight into the frequency and characteristics of deadlock in interconnection networks, nor have they identified network parameters and the degrees to which they influence deadlock [1, 2, 3, 4]. A practical approach to better understanding deadlock is through detailed simulation. We have developed an interconnection network simulator capable of true deadlock detection. We use this simulator to perform detailed deadlock characterization. This paper describes our approach for detecting ....

....approach to modeling resource dependencies, deadlock, and other types of message blocking behavior within interconnection networks. Our model supports maximally adaptive routing, and thus accommodates both restricted (deadlock avoidance based [5, 6, 7] and unrestricted (deadlock recovery based [2, 3, 4]) routing algorithms. The model is formalized in Section 3. Our resource model uses channel wait for graphs (CWGs) to represent resource allocations and requests existing within the network at a given point in time. We describe CWGs through example here and precisely define them in Section 3. The ....

[Article contains additional citation context not shown here]

Jae H. Kim, Ziqiang Liu, and Andrew A. Chien, "Compressionless Routing: A Framework for Adaptive and Fault-Tolerant Routing", In Proceedings of the 21st International Symposium on Computer Architecture, IEEE Computer Society, pages 289-300, April 1994.

....while preventing the user from pushing the network out of its operating range. While METRO requires nontrivial intelligence at the transfer endpoints, its key characteristics arise from its router design. NIFDY, in contrast, can be used with a variety of networks. Compressionless Routing (CR) KLC94] also provides in order delivery. CR, which relies on wormhole routing, pads packets with enough space to ensure that pushing the entire packet onto the network implies that the head of the packet has already entered the destination, at which point the packet is guaranteed to be completely ....

J.H. Kim, Ziqiang Liu, and A.A. Chien. Compressionless routing: a framework for adaptive and fault-tolerant routing. In Proceedings the 21st Annual International Symposium on Computer Architecture, pages 289--300. IEEE Comput. Soc. Press, 1994.

....end to end flow control, and reliable delivery, it would obviate much of the need for a messaging layer, reducing the messaging cost to the basic data movement cost. Researchers in the routing community are beginning to address such concerns; several recent projects (e.g. Compressionless Routing [17]) provide all three of these features, albeit at some cost. Communication Cost versus Latency In this study, we have focused on instruction counts as the primary measure of communication cost (performance) Latency is a reasonable alternative performance metric, but is hard to measure in a ....

J. Kim, Z. Liu, and A. Chien. Compressionless routing: A framework for adaptive and fault-tolerant routing. In Proceedings of the International Symposium on Computer Architecture, April 1994.

....services, or raise the level of services provided by the network. In the second part of this study, we examine what overheads might be recovered with routingnetworks which exploit low level hardware structure to provide higher level services. Using a network design based on Compressionless Routing [16] which provides in order delivery, end to end flow control, and packetlevel fault tolerance, we show that the software overheads of buffer management, in order delivery and end to end faulttolerance can be completely eliminated. Consequently, we propose designing networks which simplify the ....

....implementation which demonstrates that if a network provides high level services (in order delivery, end to end flow control, and reliable packet delivery) essentially all but the data movement cost can be removed from the messaging layer. Several networks with such features have been proposed [22, 16]. The remainder of the paper is organized as follows. Section 2 describes the problem context. In Section 3, we analyze the costs incurred by CMAM based implementations of typical communication protocols. Section 4 explores the advantages of higher level network features on messaging layer costs. ....

[Article contains additional citation context not shown here]

J. Kim, Z. Liu, and A. Chien. Compressionless routing: A framework for adaptive and fault-tolerant routing. In Proceedings of the International Symposium on Computer Architecture, April 1994.

....as a way to improve network performance by providing multiple paths between a source and destination. Adaptive routing techniques can produce higher utilization of network resources and more robust performance. A number of dramatically simpler adaptive routing algorithms have been proposed [11, 9, 14, 3, 12]. Such breakthroughs make adaptive routing feasible, but not without cost [4] Three significant costs remain. First, header update and selection is expensive in adaptive routers; Second, virtual channels are expensive in terms of latency and cycle time; Third, requirements of larger crossbars and ....

J. Kim, Z. Liu, and A. Chien. Compressionless routing: A framework for adaptive and fault-tolerant routing. In Proceedings of the International Symposium on Computer Architecture, April 1994.

No context found.

J. H. Kim, Z. Liu, and A. A. Chien. Compressionless Routing: A Framework for Adaptive and Fault-Tolerant Routing. In Proceedings of the 21st International Symposium on Computer Architecture, April 1994.

No context found.

J. Kim, Z. Liu and A. Chien. Compressionless Routing: A Framework for Adaptive and Fault-tolerant Routing. In Proceedings of the International Symposium on Computer Architecture, pages 289-300, 1994.

No context found.

J H Kim, Z Liu, AA Chien, "Compressionless routing: a framework for adaptive and fault-tolerant routing", Proc ISCA (Int Symposium on Computer Architecture), 1994

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