Kenneth Mackenzie, John Kubiatowicz, Matthew Frank, Walter Lee, Anant Agarwal, and M. Frans Kaashoek. UDM: User Direct Messaging for General-Purpose Multiprocessing. Technical Memo MITLCS -TM-556, MIT Laboratory for Computer Science, March 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....significantly degrade performance. I improve upon their work by exposing and examining the design space of data transfer and buffering parameters. Blumrich, et al. 14] compared the SHRIMP I and SHRIMP II NIs, but did not explore alternate data transfer and buffering mechanisms. Mackenzie, et al. [73] studied the effect of buffering using a synthetic workload and concluded that buffering messages in virtual memory can occur only rarely for realistic applications. However, in contrast I found that for two of my seven macrobenchmarks, buffering can play a significant role in improving ....

Kenneth Mackenzie, John Kubiatowicz, Matthew Frank, Walter Lee, Anant Agarwal, and M. Frans Kaashoek. UDM: User Direct Messaging for General-Purpose Multiprocessing. Technical Memo 556, MIT Laboratory for Computer Science, March 1996.

....that OS abstractions will be implemented in user space either as part of the application or, more commonly, as a separate library, a library operating system. Three implementations of exokernels are known at this time: Aegis[13] created for the MIPS R3000 processor (Digital DECstation) Glaze[27], created for the Fugu multiprocessor (MIT research Supported by Ericsson CSLAB SARC processor) and XOK[25] created for the Intel x86 processor family (Pentium, Pentium Pro and Pentium II based PC compatibles) It is our belief that for exokernels to be viable, it must be shown that ....

K. Mackenzie, J. Kubiatowicz, M. Frank, W. Lee, V. Lee, A. Agarwal, and M.F. Kaashoek. UDM: User direct messaging for general-purpose multiprocessing. Technical memo MIT/LCS/TM-556, Massachusetts Institute of Technology, March 1996.

....simple, isolated resources such as physical memory, or can it apply to more difficult shared resources such as disk Can one build an exokernel system Fortunately for this thesis, one can. There have been three exokernel systems built so far, in increasing verisimilitude: Aegis [25] Glaze [60], and Xok [48] Most of our performance data and examples come from Xok and ExOS, its default libOS. While ExOS does not handle some Unix corner cases, it is not a toy either. For example, it runs most Unix applications (e.g. perl, 2 Aegis was roughly a factor of ten faster than a mature ....

....are completely at application level, the exokernel may be unable to distinguish pages used to cache disk blocks and pages used for virtual memory. Glaze, the Fugu exokernel, has the additional complication that it cannot distinguish such uses from the physical pages used for buffering messages [60]. Frequently used information can often be derived with little effort. For example, if page tables are managed by the application, the exokernel can approximate LRU page ordering by tracking the insertion of translations into the TLB. However, at the very least, this inference requires thought. ....

Kenneth Mackenzie, John Kubiatowicz, Matthew Frank, Walter Lee, Victor Lee, Anant Agarwal, and M. Frans Kaashoek. UDM: User Direct Messaging for General-Purpose Multiprocessing. Technical Memo MIT/LCS/TM556, March 1996.

....significantly degrade performance. We improve upon their work by exposing and examining the design space of data transfer and buffering parameters. Blumrich, et al. 3] compared the SHRIMP I and SHRIMP II NIs, but did not explore alternate data transfer and buffering mechanisms. Mackenzie, et al. [25] studied the effect of buffering using a synthetic workload and concluded that buffering messages in virtual memory can occur only rarely for realistic applications. However, in contrast we found that for two of our seven macrobenchmarks, buffering can play a significant role in improving ....

Kenneth Mackenzie, John Kubiatowicz, Matthew Frank, Walter Lee, Anant Agarwal, and M. Frans Kaashoek. UDM: User Direct Messaging for GeneralPurpose Multiprocessing. Technical Memo 556, MIT Laboratory for Computer Science, March 1996.

....Active Messages [33] is general enough to implement more complex communication and synchronization protocols, which we believe makes it a good basis for algorithm analysis in modern parallel systems. A number of existing machines provide efficient support for communication with active messages [2, 9, 14, 20, 24, 28, 32]. Wemake two further assumptions in the LoPC model that appear to result in very little loss in accuracy yet great gain in simplicity. First, we assume that the hardware message buffers at the nodes are infinitely large. Second, we assume that the interconnect is contention free. We model ....

Kenneth Mackenzie, John Kubiatowicz, Matthew Frank, Walter Lee, Anant Agarwal, and M. Frans Kaashoek. UDM: User Direct Messaging for General-Purpose Multiprocessing. Technical Memo MITLCS -TM-556, MIT Laboratory for Computer Science, March 1996.

....Section 3. We provide information about the Fugu multiprocessor, the scheduler, and the multiprocessor simulator used by the experiments. Fugu is an experimental, distributed memory multiprocessor supporting both cache coherent shared memory and fine grain message passing communication mechanisms [13]. The applications studied in this paper use only the messagepassing mechanism. Messages in Fugu have extremely low overhead, costing roughly 10 cycles to send and roughly 100 cycles to process a null active message via an interrupt. The Fugu operating system, Glaze, supports virtual memory, ....

K. Mackenzie, J. Kubiatowicz, M. Frank, W. Lee, V. Lee, A. Agarwal, and M. F. Kaashoek. UDM: User Direct Messaging for General-Purpose Multiprocessing. Technical Memo MIT/LCS/TM-556, March 1996.

....mechanisms for protection and virtual memory in multiprocessors that support fast messaging. Implementing a virtual machine model is challenging because features such as multiprogramming and demand paging tend to interfere with streamlined, tightly coupled communication. Our FUGU architecture [23] embodies an optimistic approach to message passing in a virtual machine: each application has direct access to a simple, fast, Alewife like network interface unless immediate conditions (detected in hardware) dictate otherwise. Operating system software maintains the fiction of a virtual machine ....

K. Mackenzie, J. Kubiatowicz, M. Frank, W. Lee, V. Lee, A. Agarwal, and M. Frans Kaashoek, "UDM: User direct messaging for general-purpose multiprocessing," Massachusetts Inst. Technol., Cambridge, Tech. Memo MIT/LCS/TM-556, Mar. 1996.

....Active Messages [33] is general enough to implement more complex communication and synchronization protocols, which we believe makes it a good basis for algorithm analysis in modern parallel systems. A number of existing machines provide efficient support for communication with active messages [2, 9, 14, 20, 24, 28, 32]. Wemake two further assumptions in the LoPC model that appear to result in very little loss in accuracy yet great gain in simplicity. First, we assume that the hardware message buffers at the nodes are infinitely large. Second, we assume that the interconnect is contention free. We model ....

Kenneth Mackenzie, John Kubiatowicz, Matthew Frank, Walter Lee, Anant Agarwal, and M. Frans Kaashoek. UDM: User Direct Messaging for General-Purpose Multiprocessing. Technical Memo MITLCS -TM-556, MIT Laboratory for Computer Science, March 1996.

....mechanisms for protection and virtual memory in multiprocessors that support fast messaging. Implementing a virtual machine model is challenging because features such as multiprogramming and demand paging tend to interfere with streamlined, tightly coupled communication. Our FUGU architecture [23] embodies an optimistic approach to message passing in a virtual machine: each application has direct access to a simple, fast, Alewife like network interface unless immediate conditions (detected in hardware) dictate otherwise. Operating system software maintains the fiction of a virtual machine ....

Kenneth Mackenzie, John Kubiatowicz, Matthew Frank, Walter Lee, Victor Lee, Anant Agarwal, and M. Frans Kaashoek. UDM: User Direct Messaging for General-Purpose Multiprocessing. Technical Memo MIT/LCS/TM-556, March 1996.

....4 and 5. It provides information about the Fugu multiprocessor, the scheduler, and the multiprocessor simulator used by the experiments. Fugu is an experimental, distributed memory multiprocessor supporting both cache coherentshared memory and fine grain message passing communication mechanisms [12]. The applications studied in this paper use only the message passing mechanism. Messages in Fugu have extremely low overhead, costing 10 cycles to send and 100 cycles to process a null active message via an interrupt. The Fugu operating system, Glaze, supports virtual memory, preemptive ....

K. Mackenzie, J. Kubiatowicz, M. Frank, W. Lee, V. Lee, A. Agarwal, and M. F. Kaashoek. UDM: User Direct Messaging for GeneralPurpose Multiprocessing.Technical MemoMIT/LCS/TM-556,March 1996.

....XN, and explain how unprivileged applications can define new file systems and how these file systems can safely multiplex the same disk at a fine granularity. Finally, we summarize what we have learned from building three complete exokernel systems (Xok, Aegis [11] for DECstations, and Glaze [29] for the Fugu multiprocessor) The rest of the paper is organized as follows. Section 2 discusses related work. Section 3 summarizes the exokernel architecture. Section 4 provides a detailed example of reconciling application control with protection by presenting the disk system XN. Section 5 ....

....are completely at application level, the exokernel may be unable to distinguish pages used to cache disk blocks and pages used for virtual memory. Glaze, the Fugu exokernel, has the additional complication that it cannot distinguish such uses from the physical pages used for buffering messages [29]. Frequently used information can often be derived with little effort. For example, if page tables are managed by the application, the exokernel can approximate LRU page ordering by tracking the insertion of translations into the TLB. However, at the very least, this inference requires thought. ....

K. Mackenzie, J. Kubiatowicz, M. Frank, W. Lee, V. Lee, A. Agarwal, and M.F. Kaashoek. UDM: user direct messaging for general-purpose multiprocessing. Technical Memo MIT/LCS/TM-556, March 1996.

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