Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A substrate for portable parallel programs. In COMPCON '95, pages 327-332, San Francisco, California, March 1995. IEEE Computer Society.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....system, this mechanism is trivially implemented using system calls that change the page protection for the shared heap. detects reads or writes to invalid blocks or writes to read only blocks and traps to user level code, which uses the resulting exceptions to execute a coherence protocol action [HLW95] At each memory reference, the system must ensure that the referenced datum is local and accessible. Therefore, each memory reference is logically preceded by a check that has the following if ( lookup (Address) protocol software must be invoked. This sequence must execute atomically to ....

....coherence protocols to be implemented as user level libraries. Tempest s fine grain access control mechanism detects reads or writes to invalid blocks or writes to read only blocks and traps to user level code, which uses the resulting exceptions to execute a coherence protocol action in software [HLW95] At each memory reference, the system must ensure that the referenced datum is local and accessible. If the access is allowed, the reference proceeds normally. Otherwise, the shared memory protocol software must be invoked, which is achieved by suspending the computation and invoking a ....

[Article contains additional citation context not shown here]

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A substrate for portable parallel programs. In COMPCON '95, pages 327-332, San Francisco, California, March 1995. IEEE Computer Society.

....it in the network interface. These schemes require the compiler or the linker to automatically convert remote loads and stores in parallel programs into other communication primitives, such as send and receive primitives. Some examples are Ivy [51] Munin [52] Midway [53] Tempest and Blizzard [54], and TreadMarks [55] These schemes help improve the portability of programs across parallel systems, regardless of the organization of the network interface. For some programs, the performance of these software schemes approach that of native hardware support for remote memory. While remote ....

Mark D. Hill, James R. Larus and David A. Wood. Tempest: a substrate for portable parallel programs. Proc. of COMPCON, Mar. 1995.

....We model shared memory systems by having a collection of ReadWrite objects. The abstract state of each such object consists of an integer or machine word. Other kinds of systems can be modeled as well, including hybrids of the message passing and shared memory paradigms, such as those described in [8, 10, 11, 15, 19]. The proposed research is to explore the theory and practice of this objectoriented model of concurrent systems, and to investigate how results in the general model apply to the subdomains of message passing and shared memory. The first step of this research is to elaborate the model. The second ....

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A substrate for portable parallel programs. In COMPCON '95, pages 327-- 33, San Francisco, CA, USA, 5--9 March 1995. IEEE Computer Society, IEEE Computer Society Press.

....which exemplify the debate of policies and mechanisms. The Scalable Coherent Interface (SCI) coherence protocol [12] is a hardware based policy for communicatiing in large scale multiprocessors. Typhoon [1] is a proposed high end (hardware assisted) implementation of the Tempest Interface [3], a portable parallel programming interface which provides a diverse set of mechanisms for communicating in large scale multiprocessors. The sections which follow briefly summarize the two approaches. 1.1 SCI SCI is a distributed directory based hardware coherency protocol. The objective of SCI ....

....is no single coherence protocol that works optimally over a broad range of applications. Therefore, an architecture should provide mechanisms instead of policies (protocols) to the user for handling communication. This leads to the concepts of user level shared memory and the Tempest Interface [3]. ffl Certain communication namely, fine grained sharing in the form of coherence messages requires the ability to communicate at the cache block level. This leads to the concept of hardware assisted fine grained access control. ffl Since it is sometimes possible to overlap communication ....

Mark D. Hill, James R. Larus, and David A. Wood, "Tempest: A Substrate for Portable Parallel Programs, " COMPCON Spring 1995, March 1995.

....to end users. The range of applications being considered for development on FM, or already in the works, is testimonial of the flexibility of its interface. Besides the MPI message passing library, these are the BSD socket interface, the Converse compiler backend [15] the Tempest runtime library [12], the Orca Project parallel object language [3] FM performance exceeds the performance of messaging layers presently available on a number of MPPs [14, 24, 8] FM 1.1 is the version which has been used to build the MPI library upon. The two graphs of figure 2.1 show the performance of FM. The ....

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A substrate for portable parallel programs. In Compcon, March 1995. Available from ftp://ftp.cs.wisc.edu/wwt/compcon95 tempest.ps.

....all simulations in this chapter, I assume that the number of network message buffers allocated at the sender and receiver respectively is fixed at eight. Chapter 5 examines the effect of varying this parameter on several NIs. I ran all my benchmarks on the Tempest parallel programming interface [52]. Messagepassing benchmarks use only Tempest s active messages. Shared memory codes on Tempest also use active messages, but assume hardware support for fine grain access control [103] Codes with custom protocols use a combination of the two. 4.2.3 Macrobenchmarks I use seven macrobenchmarks ....

....Wisconsin Stache Protocol I obtained my coherence message traces from the Wisconsin Stache protocol [100] Stache is a software, full map, and write invalidate directory protocol that uses part of local memory as a cache for remote data. Currently, Stache is implemented on the Tempest interface [52], which is a portable interface for writing shared memory programs. Table 6.1 shows all the types of coherence messages generated by Stache. These coherence messages are also common to most full map, write invalidate directory protocols. For all my simulations with Stache I use a (software) cache ....

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A Substrate for Portable Parallel Programs. In COMPCON '95, pages 327--332, San Francisco, California, March 1995. IEEE Computer Society.

.... [2] Tempest aims to provide a standard, system independent parallel programming user system interface that offers programmers access to a variety of different communication mechanisms, including active messages, bulk data transfer, virtual memory management, and fine grain access control [3]. Tempest defines the architecture of a communication interface for shared memory parallel programs; Blizzard is one implementation of that architecture [4] Blizzard runs the coherence protocol code in software on each of the main compute nodes. Studies with Blizzard on applications with a wide ....

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A Substrate for Portable Parallel Programs. COMP/CON Spring 95.

....across both platforms and programming models. As also discussed in [20] parallel hardware is converging towards architectures in which compute nodes interconnected via network interfaces like Myrinet are common. Our monitoring code would work on any system currently using the Myrinet interface [2, 13, 23]; we also expect that porting to similar systems with other programmable interface processors would be fairly straightforward. To gather low level statistics, our approach represents a significant improvement over previous reliance on simulation techniques or custom designed hardware monitor ....

M. Hill, J. R. Larus, and D. A. Wood. Tempest: A Substrate for Portable Parallel Programs. In Proc. COMPCON, Mar 1995.

....a shared object abstraction on workstation clusters for executing parallel programs. Some examples are IVY [16] Amber [10] Munin [9] Orca [4] and Treadmarks [15] Current research in using workstation clusters to run parallel applications includes work at Princeton [6] University of Wisconsin [13], University of Washington [24] Berkeley [1] and Cornell [26] In the past ten years, there has also been con siderable effort in designing hardware and software interfaces to minimize the cost of network access [7, 12, 14, 19, 23] Our primary goal was to build a working system of reasonable ....

M. D. Hill, J. R. Larus, and D. A. Wood. Tempest: A substrate for portable parallel programs. In COMPCON '95, Mar. 1995.

....to program both shared and distributed memory machines in a similar fashion. Demands for increased programmability of distributed memory machines have given rise to research efforts resulting in Distributed Shared Memory (DSM) libraries layered on top of the networking and virtual memory systems [6,7,14,19,23]. They have also resulted in the development of distributed object oriented systems offering user programs network wide access to shared services [12,18,32] as well as object oriented concurrent programming layers on top of shared and distributed memory machines [10,21,33,36] Our contribution to ....

....with the lower RPC like layers of distributed object systems like Spring [32] Chorus [27] and distributed objects [33] Indigo attempts to combine both the functionality required by distributed objects and by distributed shared memory. Indigo s goals are similar to those of the Tempest library [19] but it differs from Tempest in its additional support for objects and in its exclusion of constructs required only for DSM implementations (e.g. read write access detection) Such DSM specific support is implemented above Indigo s level of abstraction. Our hope is that libraries like Indigo may ....

[Article contains additional citation context not shown here]

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A substrate for portable parallel programs. In COMPCON, March 1995.

....Blizzard runs under an unmodified Solaris 2.4 kernel. Blizzard also supports more efficient fine grain access control, including optimized software and dedicated hardware. Again, the final difference is that Blizzard exploits multiprocessor nodes. Blizzard s uses Tempest active messages [15,38,27]. Tempest communication differs from other active message system [13,25,37] since it does not constrain a program to request response protocols and interrupts computation on message arrival. These unique aspects are necessary for coherence protocols, but complicate the communication layer. Even ....

....to the kernel address space. 4 Fine Grain Access Control Fine grain access control is the Tempest mechanism that detects reads or writes to invalid blocks or writes to read only blocks and traps to user level code, which uses the resulting exceptions to execute a coherence protocol action [15]. We implemented three techniques for fine grain access control on commodity systems. 4.1 Blizzard S: Software Access Control Blizzard S uses the EEL executable editing library [20] to insert access control tests before shared memory loads and stores in a fully compiled and linked program. ....

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A Substrate for Portable Parallel Programs. In COMPCON '95, pages 327--332, San Francisco, California, March 1995. IEEE Computer Society.

....as bus snooping and memory management. SHRIMP s hardware implementation overhead is higher than FUNi s, and SHRIMP s association with a specific processor memory bus limits its upgrade path. SHRIMP also does not have the benefit of a network coprocessor. Several academic and industry NOW projects[1, 7, 11, 15, 16, 18] have based their research on a cluster of workstations interconnected by Myrinet[3] Derived from Caltech s ATOMIC project[6] Myrinet is available commercially as a ready to go high performance NOW interconnection package. The Myrinet package comes complete with network routers, end point ....

M. D. Hill, J. R. Larus, and D. A. Wood. Tempest: A substrate for portable parallel programs. In Proceedings of COMPCON Spring 95, March 1995.

....this implementation did not generate uniformly positive results. While the protocol worked well for applications such as Water and String with regular, repetitive communication patterns, it degraded the performance of other applications by generating an excessive amount of communication. Tempest [8, 12] is a collection of communication mechanisms designed for implementation on a range of architectures. Programmers can apply communication optimizations by selecting an appropriate communication mechanism from a library or by coding their own custom protocol. Munin [5] is a page based distributed ....

M. Hill, J. Larus, and D. Wood. Tempest: A substrate for portable parallel programs. In Proceedings of the 1995 Spring COMPCON, March 1995.

.... Cycles 7 locks, if a processor blocks on a message send, it reads messages from the NI and buffers them in user space (except for CNI 16 Q m in which messages automatically overflow to main memory from the CNI cache) All benchmarks are run on the Tempest parallel programming interface [22]. Message passing benchmarks use only Tempest s active messages [37] Shared memory codes on Tempest also use active messages, but assume hardware support for fine grain access control [39] Codes with custom protocols use a combination of the two. 4.2 Macrobenchmarks Table 3 depicts five ....

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A Substrate for Portable Parallel Programs. In COMPCON '95, pages 327--332, San Francisco, California, March 1995. IEEE Computer Society.

....We model shared memory systems by having a collection of ReadWrite objects. The abstract state of each such object consists of an integer or machine word. Other kinds of systems can also be modeled, including hybrids of the message passing and shared memory paradigms such as those described in [42, 50, 54, 68, 75]. The proposed research is to investigate the use of compositional techniques in constructing transparently fault tolerant distributed object systems and proving properties of objects in such a system. In particular, the following points will be addressed: ffl Compositional reasoning techniques ....

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A substrate for portable parallel programs. In COMPCON '95, pages 32733, San Francisco, CA, USA, 59 March 1995. IEEE Computer Society, IEEE Computer Society Press.

....Wisconsin Stache Protocol We obtained our coherence message traces from the Wisconsin Stache protocol [30] Stache is a software, fullmap, and write invalidate directory protocol that uses part of local memory as a cache for remote data. Currently, Stache is implemented on the Tempest interface [16], which is a portable interface for writing shared memory programs. Table 1 shows all the types of coherence messages generated by Stache. These coherence messages are also common to most full map, write invalidate directory protocols. For all our simulations with Stache we use a (software) cache ....

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A Substrate for Portable Parallel Programs. In COMPCON '95, pages 327--332, San Francisco, California, March 1995. IEEE Computer Society.

....less flexibility in selecting levels of performance monitoring detail. Over time, scalable parallel machine designs are converging, as are the communication mechanisms in both message passing and shared memory machines. Although not hardware cache coherent, machines like the Wisconsin COW [HLW95] or upcoming SHRIMP [BLA 94] implementations may also be amenable to integrating communication and performance monitoring. In these machines, the network interface is (or is likely to be) implemented using Myricom switches [BCF 95] The core of these switches is a programmable controller ....

M. D. Hill, J. R. Larus, and D. A. Wood. Tempest: A Substrate for Portable Parallel Programs. Proc. Compcon. March, 1995.

....Poor responsiveness degrades performance by increasing the time a processor idles awaiting completion of a coherence operation. Other DSM techniques [8, 15] which optimize synchronous protocols by predicting object access patterns are less effective given the dynamic context. Recent systems [3, 11] which allow application customization of coherence protocols provide mechanisms for improving performance, but offer little insight into how to build protocols for tolerating coherence processing delays arising from unresponsive processors. We describe a new caching framework View Caching ....

....operations from the critical path and replace synchronous protocols with asynchronous ones. However, such optimizations are less effective in our dynamic context where predictive information about future object accesses is hard to obtain. More promising are systems such as Munin [3] and Tempest [11] which allow application specific customization of coherence protocols. Munin allows custom coherence actions for shared regions by selecting from among a set of predefined coherence protocols that reflect common application sharing scenarios. While such an approach improves performance in certain ....

[Article contains additional citation context not shown here]

M. D. Hill, J. R. Larus, and D. A. Wood. Tempest: A substrate for portable parallel programs. In Proceedings of COMPCON, Mar. 1995.

....cached for subsequent references) and to maintain the global state necessary to implement a coherent programming model. At a load or store, the actual communication depends on the memory system s protocols and the program s dynamic state in ways that can be difficult to understand or predict [9]. Nevertheless, detecting this hidden communication is essential to isolate and eliminate many shared memory performance bottlenecks. For example, a cache block s migratory behavior may indicate false sharing, which a programmer can eliminate by aligning and padding data structures. Moreover, ....

M. D. Hill, J.R. Larus, S.K. Reinhardt, and D.A. Wood. Tempest: A Substrate for Portable Parallel Programs. COMPCON'95, San Francisco, March 1995.

....subset of these target architectures will be made available in the final distribution of WWT II. Differences WWT WWT II Host Platforms TMC CM 5 Workstation, SMP, COW, COW SMP Target Architectures CC NUMA CC NUMA, S COMA, software DSM, SMP, and Tempest (active messages and shared memory)[11] a Memory Bus Contentionfree Detailed simulation of a coherent memory bus Network Optional Network Simulation Network is not modeled, but network contention is modeled at the network interfaces Source Language C C (primarily) Number of non blank, non comment lines of ....

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A Substrate for Portable Parallel Programs. In COMPCON '95, pages 327--332, San Francisco, California, March 1995. IEEE Computer Society.

....from) the SAN interface. LANs use very heavy weight software protocols such as TCP IP, which make the conservative assumption that LANs, like the internet, are highly unreliable. For SANs, such overly conservative protocols can be replaced with lean communication layers such as Active Messages [54, 24]. Typically, LAN interfaces deliver messages to user applications via the operating system, which can be very expensive. For SANs, the latency through the operating system can be eliminated by providing applications with direct user level access to the network interface hardware. For example, the ....

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A Substrate for Portable Parallel Programs. In COMPCON '95, pages 327--332, San Francisco, California, March 1995. IEEE Computer Society.

....to these processors. This is because our primary focus is on relative performance of different NIs using the same base processor model, and not on the absolute performance of a particular processor architecture. All of our benchmarks are run on the Tempest parallel programming interface [19]. Message passing benchmarks use only Tempest s active messages. Shared memory codes on Tempest also use active messages, but assume hardware support for fine grain access control. Codes with custom protocols use a combination of the two. 5.1.2 Network and Flow Control All of our simulations ....

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A Substrate for Portable Parallel Programs. In COMPCON '95, pages 327--332, San Francisco, California, March 1995. IEEE Computer Society.

No context found.

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A Substrate for Portable Parallel Programs. In COMPCON '95, pages 327--332, San Francisco, California, March 1995. IEEE Computer Society.

.... another alternative: hybrid shared memory and message passing machines that offer programmers the opportunity to select coherence protocols and fall back to message passing communication [13, 14, 19] The Wisconsin Wind Tunnel project s approach is a portable, user level interface called Tempest [11, 19, 18], which provides message passing communication and mechanisms to construct shared memory protocols. In particular, Tempest provides programs with the novel ability to copy and move data without changing its address (renaming it) In this paper, we use an implementation of Tempest called Blizzard ....

....shared memory, and other hybrid models. Tempest is designed so that it can be supported on many platforms, providing portability across these systems. The Blizzard system implements the Tempest substrate on a Thinking Machines CM 5 and is being ported to the Wisconsin COW (Cluster of Workstations) [11]. for (i = 0; i number timesteps; i ) f . for (j = start edge; j = end edge; j ) f n1 = edge[j] left node; n2 = edge[j] right node; w = f(n1, n2, j) lock(y lock[n1] y[n1] g1(x[n1] x[n2] w) unlock(y lock[n1] lock(y lock[n2] y[n2] g2(x[n1] x[n2] w) unlock(y ....

[Article contains additional citation context not shown here]

Mark D. Hill, James R. Larus, and David A. Wood. Tempest: A Substrate for Portable Parallel Programs. In COMPCON '95, pages 327--332, San Francisco, California, March 1995. IEEE Computer Society.

No context found.

Mark D. Hill, James R. Larus, David A. Wood, "Tempest: A Substrate for Portable Parallel Programs." COMPCON Spring 95, 1995.

First 50 documents

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