Kenneth Mackenzie, John Kubiatowicz, Matthew Frank, Walter Lee, Victor Lee, Anant Agarwal, and M. Frans Kaashoek. Exploiting Two-Case Delivery for Fast Protected Messaging. In Proceedings of the 4th International Symposium on HighPerformance Computer Architecture, pages 231--242, February 1998.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....IOLite [23] change the user kernel interface and programming model to pass around collections of pointers. The user is aware that her data is split into various memory regions which complicates program ming. Another approach has user handlers manage the copy from the network interface directly [20]. Direct access to the network interface requires special hardware, does not interact well with multiprogramming and demand paging, and results in the entire packet, not just the payload, being transfered to user space. A final approach [9] uses page remapping, which can be implemented under the ....

K. Mackenzie, J. Kubiatowicz, M. Frank, W. Lee, V. Lee, A. Agarwal, and M. F. Kaashoek. Exploiting two-case delivery for fast protected messaging. In HPCA, pages 231--242, 1998.

....major sources of the communication overhead. The communication hardware aspect includes the architecture and placement of the network interface, and the interconnection network and its services. Many architectures have been proposed for the network interfaces. They are classified as (1) direct [52, 7, 63, 80, 97, 88] and (2) memory based [48, 112, 126, 23] Direct network interfaces allow a processor to directly access the network queue. However, they mostly ignore the issue of multiprogramming. That is, a single thread can only use the network interface at a time. Memory based interfaces provide protection ....

K. Mackenzie, J. Kubiatowiz, M. Frank, W. Lee, V. Lee, A. Agarwal and M. F. Kaashock, "Exploiting Two-Case Delivery for Fast Protected Messaging", Proceedings of the 4th International Symposium on High-Performance Computer Architecture, February 1998.

....In cases where many connections need to be created and destroyed, this can become a bottleneck, and no library seems to solve this problem in a satisfactory way for the general case. Interrupt Handling: The cost of delivering interrupts has been revealed to be a major problem in many cases [4, 14]. All libraries under consideration try to avoid expensive interrupts by having the network interface place messages directly in user memory. Furthermore, libraries that require handlers to run at receipt of each message, specifically AM, avoid up calls by polling for network events at the ....

K. Mackenzie, J. Kubiatowicz, M. Frank, W. Lee, V. Lee, A. Agarwal, and M. F. Kaashoek. Exploiting two-case delivery for fast protected messaging. In The 4th IEEE Symposium on High-Performance Computer Architecture, Feb 1998.

....with these links to create parallel systems of arbitrary topology. They also allow a VIRAM chip to be used along with a hard disk drive as the building block for scalable data servers like ISTORE [7] The structure of the network interface is similar to those in the Alewife [24] and Fugu [26] systems. It is memory mapped as a virtual resource and allows applications to send short or long messages without invoking the operating system. Short messages can be created by storing data directly into a message buffer in the network interface. For long messages, one or more DMA descriptors ....

K. Mackenzie, J. Kubiatowicz, M. Frank, W. Lee, V. Lee, A. Agarwal, and M Kaashock. Exploiting two-case delivery for fast protected messaging. In Proceedings of the Fourth International Symposium on High-Performance Computer Architecture, pages 231--42, February 1998.

....that the destination process is not scheduled increases with more competing processes. Second, if a destination process is not scheduled, then the time remaining before that process is scheduled increases. In summary, spin waiting interacts very poorly with processes that are locally scheduled [7, 35, 55, 99, 109]. When the scheduling of processes is not coordinated, processes often do not receive the response from the remote node until their next time slice. As a result, the process spins idly for the remainder of its time slice without making progress. With spin waiting, communication often completes in ....

....and more messages may arrive in the future. The actions described above are taken regardless of whether the remote process actually requires an immediate response to make forward progress. In fact, there are numerous situations where a sending process can make forward progress without a response [46, 95, 109]; for example, the request could be a one way message storing data, or the first segment of bulk transfer, or a notification that this process has reached a barrier before other processes have done so. However, when a request message arrives, there is no way for the local process to know whether ....

Kenneth Mackenzie, John Kubiatowicz, Matthew Frank, Walter Lee, Victor Lee, Anant Agarwal, and M. Frans Kaashoek. Exploiting Two-Case Delivery for Fast Protected Messaging. In Proceedings of the 4th International Symposium on HighPerformance Computer Architecture, pages 231--242, February 1998. 202

....interface unit, to interrupt Sparcle and have software take over some of its functions. For instance, this is used to handle corner cases in Alewife s cache coherence protocol [17] Alewife does not deal with issues of protection and sharing. These are investigated in a follow on project, FUGU [70, 69, 71]. Aside from adding small hardware extensions to Alewife, FUGU relies on Sparcle s fast interrupt and interrupt software to impose sharing protection. The constraints we faced are very different from those faced in Alewife and FUGU. For example, Sparcle has no on chip cache, and Alewife s CMMU ....

K. Mackenzie, J. Kubiatowicz, M. Frank, V. Lee, A. Agarwal, and F. Kaashoek. Exploiting Two-Case Delivery for Fast Protected Messaging. In Proceedings of the Fourth International Symposium on High-Performance Computer Architecture, Feb. 1998.

....thread only if their process IDs match up. Otherwise the message is handed to the operating system, which is notified via an interrupt. To avoid starvation due to interrupt masking, the technique is further extended with a revocable interrupt disable mechanism in the FUGU architecture [30]. Nonetheless, good performance in this approach relies on an GID match being the common case, making it suited mostly for a gang scheduled machine. Note also that software libraries are written for several systems in Table 2.1 to provide a standardized messaging interface, such as MPI [31] and NX ....

....workarounds are necessary in these designs. The CM 5 typically relies on a gangscheduling scheme, which time slices the machine and drains the network at the end of every interval, while the J Machine merely counts on the user to release critical resources promptly. The more robust FUGU system [30] prevents message interception in its user level network interface by identifying each message and thread with a hardware stamp. A message is presented to the current process at the destination only if their stamps match up, and is referred to the operating system otherwise. To prevent starvation, ....

Kenneth Mackenzie, John Kubiatowicz, Matthew Frank, Walter Lee, Victor Lee, Anant Agarwal, M. Frans Kasshoek, "Exploiting Two-Case Delivery for Fast Protected Messaging", in HPCA 1998. pp. 231--242.

....combinations of mechanisms. For example, machines such as the BBN Butterfly have long supported shared memory and bulk transfer, the Cray T3E [27] supports both shared memory and messaging styles of communication, the Stanford Dash [18] supports shared memory and prefetching, MIT Alewife [1] Fugu [20], and the Wisconsin Typhoon [25] support several variants of shared memory and messaging styles. The availability of machines with multiple mechanisms has led to an increasing amount of insight on the effectiveness of the various mechanisms for different applications [6] 11] 28] 30] 15] 8] ....

K. Mackenzie et al. Exploiting two-case delivery for fast protected messaging. In HPCA-4, 1998.

....the destination process is not scheduled increases with more competing processes. Second, if a destination process is not scheduled, then the time remaining before that process is scheduled increases. 30 In summary, spin waiting interacts very poorly with processes that are locally scheduled [6, 32, 50, 95, 105]. When the scheduling of processes is not coordinated, processes often do not receive the response from the remote node until their next time slice. As a result, the process spins idly for the remainder of its time slice without making progress. With spin waiting, communication often completes in ....

....may arrive in the future. The actions described above are taken regardless of whether the remote process actually requires an immediate response in order for it to make forward progress. In fact, there are numerous situations where a sending process can make forward progress without a response [42, 91, 105]; for example, the request could be a one way message storing data, or the first segment of bulk transfer, or a notification that this process has reached a barrier before other processes have done so. However, when a request message arrives, there is no way for the local process to know whether ....

Kenneth Mackenzie, John Kubiatowicz, Matthew Frank, Walter Lee, Victor Lee, Anant Agarwal, and M. Frans Kaashoek. Exploiting Two-Case Delivery for Fast Protected Messaging. In Proceedings of the 4th International Symposium on HighPerformance Computer Architecture, pages 231--242, February 1998.

....traditional message interfaces, high latency and processor occupancy (unavailability of processing resources for productive computation during message operations) inhibit efficient exploitation of large scale parallelism. Recent designs address this problem by removing OS layers from the interface [4, 8, 10, 11, 12]. Nonetheless, the remaining overhead is still large. For instance, the initialization cost alone for a round trip null message is 330 cycles in the SHRIMP [12] To amortize communication overhead of 100s of cycles, programmers are motivated to use messages that are 100s to 1000s of words or ....

....operations and appropriate flush instructions. To prevent message interception, the CM 5 [4] drains the network at the end of each time slice, during which only threads from one same job are scheduled. Such gang scheduling techniques have very high global synchronization overhead. The FUGU system [10] uses a hardware stamp to identify each message and thread. A message is presented to its receiver only if the correct receiver is currently installed. Otherwise, the operating system takes over. An analogous but insecure software based tagging system is used in the revised Active Message ....

Kenneth Mackenzie et. al, "Exploiting Two-Case Delivery for Fast Protected Messaging." in HPCA 1998. pp 231--242.

No context found.

MACKENZIE, KUBIATOWICZ, FRANK, LEE, LEE, AGARWAL, AND KAASHOEK. Exploiting Two-Case Delivery for Fast Protected Messaging. In HPCA (July 1997).

....immediately. Demandpaged virtual memory causes a similar effects: a page fault (or a remote shared memory miss) in message handling code introduces a delay in message reception that may be intolerable to the network. We proposed a solution to the VNI problem in [50] and partly evaluated it in [51]. This VNI solution is implemented in FUGU and is the focus of this thesis. Other recent network interface work addresses the VNI problem with similar goals, notably CNI [58] the T family [61, 2] and the M machine [25] These projects are described as related work in Chapter 8. ffl Second is ....

....buffer will clear relatively quickly. 5 The first is that any application that requires a reply after message send inherently limits its own communication rate. Limiting the number of outstanding requests guarantees that, if buffering 5 This section is largely the work of Matt Frank as part of [51] # synth 1000 # synth 100 # synth 10 0 500 1000 1500 2000 0 10 20 Instructions between message sends Percent of messages buffered # # # # # # # # # # # # # # # # # # ## # # # # # # # Figure 7 21. Fraction of messages buffered versus send interval (T interhandler ) with N messages ....

Kenneth Mackenzie, John Kubiatowicz, Matthew Frank, Walter Lee, Victor Lee, Anant Agarwal, and M. Frans Kaashoek. Exploiting Two-Case Delivery for Fast Protected Messaging. In Proceedings of the Fourth International Symposium on High-Performance Computer Architecture, February 1998.

....a new cost model, called LoGPC, that extends the LogP [9] and LogGP [4] models to account for the impact of network contention and network interface DMA behavior on the performance of message passing programs. We validate LoGPC by analyzing three applications implemented with Active Messages [11, 18] on the MIT Alewife multiprocessor. Our analysis shows that network contention accounts for up to 50 of the total execution time. In addition, we show that the impact of communication locality on the communication costs is at most a factor of two on Alewife. Finally, we use the model to identify ....

....on the message injection rate. Finally, LoGPC models the pipelining characteristics of DMA engines which allow the overlap of memory and network access times. We validate LoGPC by comparing its predictions to the measured performance of three applications implemented with Active Messages [11, 18] on the MIT Alewife [2] multiprocessor. The applications used for validation are all to all remap with synchronous and asynchronous messaging, a dynamic programming based DNA chain comparison program called the Diamond DAG, and EM3D a benchmark code that models the propagation of the ....

K. Mackenzie, J. Kubiatowicz, M. Frank, W. Lee, V. Lee, A. Agarwal, and F. Kaashoek, "Exploiting Two-Case Delivery for Fast Protected Messaging," Proc. of 4th Int'l Symposium on High Performance Computer Architecture, Las Vegas, NV, Feb. 1998.

No context found.

Kenneth Mackenzie, John Kubiatowicz, Matthew Frank, Walter Lee, Victor Lee, Anant Agarwal, and M. Frans Kaashoek. Exploiting Two-Case Delivery for Fast Protected Messaging. In Proceedings of the 4th International Symposium on HighPerformance Computer Architecture, pages 231--242, February 1998.

No context found.

K. Mackenzie et al. Exploiting Two-Case Delivery for Fast Protected Messaging. In Proceedings of the Fourth IEEE Symposium on High-Performance Computer Architecture, Feb. 1998.

No context found.

K. Mackenzie et al., "Exploiting Two-Case Delivery for Fast Protected Messaging," Proc. High-Performance Computer Architecture, IEEE CS Press, Los Alamitos, Calif., 1998, pp. 231-242.

No context found.

K. Mackenzie et al. "Exploiting Two-Case Delivery for Fast Protected Messaging." Proceedings of 4th International Symposium on High Performance Computer Architecture Feb. 1998.

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