R. Unrau, O. Krieger, B. Gamsa, M. Stumm. Hierarchical Clustering: A Structure for Scalable Multiprocessor Operating System Design. Journal of Supercomputing, 1995.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....locks that need to be acquired. Some building blocks, such as those used for a shared file or the Process building block used for a parallel program, may be widely shared across a large multiprocessor. To implement these widely shared building blocks efficiently, the concept of clustered objects [1, 12, 29] has been developed. A clustered object building block is one that can be partitioned into representative (rep) objects, where independent requests on different processors are (in the common case) handled by different representatives of the object. A clustered object is like any other building ....

....K42 Operating System K42 is an operating system designed from the ground up and targeted at multiprocessor shared memory machines. The project is being conducted at IBM Watson Research with collaboration from University of Toronto and other universities. K42 was heavily influenced by Tornado [1, 2, 11, 12, 29], developed at the University of Toronto. Achieving good performance for shared memory multiprocessor programs has received considerable attention [5, 12, 15, 19, 29] K42 s overall structure, algorithms, and data structures have been designed with the purpose of achieving good multiprocessor ....

[Article contains additional citation context not shown here]

R.C. Unrau, O. Krieger, B. Gamsa, and M. Stumm. Hierarchical clustering: A structure for scalable multiprocessor operating system design. Journal of Supercomputing, 9(1/2):105-- 134, 1995.

....are then developed and implemented, resulting in an optimized and scalable le system. 1. 2 Hurricane File System The Hurricane File System was originally designed and implemented by Orran Krieger [34] at the University of Toronto for the Hector Multiprocessor [83] and Hurricane Operating System [81]. HFS was speci cally designed for large scale, shared memory, NUMA multiprocessor computers. The main design goals of HFS were exibility and scalability. Due to the common objectives between HFS and K42, HFS contains a valuable framework to build upon. HFS di ers from other le systems in a ....

Ron Unrau, Michael Stumm, Orran Krieger, and Ben Gamsa. Hierarchical clustering: A structure for scalable multiprocessor operating system design. Journal of Supercomputing, 1995.

....Processor pools are an operating system construct for scheduling parallel applications. The system is partitioned into a fixed set of equal sized pools, and the threads of a parallel job are scheduled to run within a single processor pool to improve memory locality. The Hurricane operating system [70] uses a hierarchical clustering approach to improve performance and scalability in NUMA multiprocessors. Hurricane partitions hardware and software resources into clusters. Operating system resources are partitioned and replicated across the clusters to reduce resource contention within the ....

R. Unrau, O. Krieger, B. Gamsa, and M. Stumm. Hierarchical clustering: A structure for scalable multiprocessor operating system design. Journal of Supercomputing, 9:105--134, 1995.

....The operating system provides 2 level scheduling policies by partitioning and partition internal thread queues. Hector[32] connects clusters of processing modules by hierarchical ring network. Hurricane on Hector provides processor pool based scheduling[40, 4] Its Hierarchical Clustering [35] scheme enables to schedule threads of an application close to each other with granularity of each level of operating system structure hierarchy. However, the threads are not gang scheduled and no resource related informations are used. Affinity scheduling[36, 31] mostly used for UMA systems ....

R. Unrau, O. Krieger, B. Gamsa, and M. Stumm. Hierarchical Clustering: A Structure for Scalable Multiprocessor Operating System Design. USENIX, 1992.

....be added without affecting the overall performance of the system. Considering that the Web is our ultimate system, no bottleneck is allowed. Figure 1: WebOSys architecture CASCON99 WebOSys Page: 6 Other distributed systems have similar requirements. Best suited for our design are Hurricane [UKGS95] and GLUnix [GPR 97] That is, we use hierarchical clusters like in Hurricane and each cluster has a centralized architecture with a decision component like in GLUnix. Three different participants, brokers, hosts and clients compose our clusters. The broker is the central decision component and ....

Ron Urau, Orran Krieger, Benjamin Gamsa, Micheal Stumm, "Hierarchical Clustering: A structure for Scalable Multiprocessor Operating System Design", Journal of Supercomputing, 1995.

....workload that includes an outof core application. Although we introduced the concept of release operations in that earlier paper, we made little use of them because they offered no significant performance benefit to stand alone out of core applications on the research prototype OS (Hurricane [21]) and machine (Hector [22] that we used. Note that we observe a different result in this study using a modern commercial OS and machine. The primary contribution of this paper is that we propose, implement, and evaluate a solution to the problem of preventing out of core applications from ....

R. C. Unrau, O. Krieger, B. Gamsa, and M. Stumm. Hierarchical Clustering: A Structure for Scalable Multiprocessor Operating System Design. Journal of Supercomputing, 9(1/2):105--134, 1995.

....system end architectures in directions relevant to future computational grids. To list just three: Operating systems are evolving to support operation in clustered environments, in which services are distributed over multiple networked computers, rather than replicated on every processor [3] [65]. A second important trend is toward a greater integration of end systems (computers, disks, etc. with networks, with the goal of reducing the overheads incurred at network interfaces and hence increasing communication rates [22] 35] Finally, support for mobile code is starting to appear, in ....

....for allocating resources, creating processes, controlling processes, accessing files, and so forth, that work regardless of a program s location within the cluster. However, when performance is critical, new implementation techniques, low level services, and high level interfaces can be required [65], 25] Future Directions Cluster architectures are evolving in response to three pressures: 1. Performance requirements motivate increased integration and hence operating system and hardware modifications (for example, to support fast communications) 2. Changed operational parameters introduce ....

R. Unrau, O. Krieger, B. Gamsa, and M. Stumm. Hierarchical clustering: A structure for scalable multiprocessor operating system design. The Journal of Supercomputing, 9(1/2):105--134, 1995.

....nodes and support NUMA systems [IRIX00] Unfortunately, such an overhaul usually requires a significant software development effort and will not support fault containment, since the base operating system is inherently vulnerable to faults. Newer operating system developments, such as the Tornado [Unrau95] and K42 [K4200] projects, have been proposed to address scalability. While these approaches tackle the problem at the basic level, they require a very significant development time and cost before reaching commercial maturity. Distributed operating systems Various forms of fault containment has ....

R.C. Unrau, O. Krieger, B. Gamsa, and M. Stumm. "Hierarchical clustering: A structure for scalable multiprocessor operating system design." Journal of Supercomputing, vol. 9, no. 1, pp.105-134, 1995.

....including the hardware platform and the software infrastructure used, and the applications which we study in our experiments. 4.1. 1 Hardware and Software Infrastructure The experimental platform used to evaluate our scheme is the Hurricane File System (HFS) 18] and Hurricane operating system [34] running on the Hector sharedmemory multiprocessor [35] Hurricane is a hierarchically clustered, micro kernel based 43 Chapter 4. Evaluation of Proposed System 44 Table 4.1: Experimental platform characteristics. Processor Processor type: Motorola 88100 Clock rate: 16.67 MHz Data cache size: ....

....file (for example, for replicated files, the application can specify which replica should be used) The basic characteristics of our experimental platform (with the instrumentation disabled) are shown in Table 4. 1, and more detailed descriptions of the platform can be found in earlier publications [18, 34, 35] Our experiments are performed on a 16 processor Hector prototype with seven Conner CP3200 disks attached to it. Each disk is directly attached to a different processor Chapter 4. Evaluation of Proposed System 45 and the local processor is responsible for initiating all requests to its disk and ....

R. C. Unrau, O. Krieger, B. Gamsa, and M. Stumm. Hierarchical clustering: A structure for scalable multiprocessor operating system design. Journal of Supercomputing, 9(1/2):105--134, 1995.

....scalability. Processes running on separate cells share few kernel resources, so operating system parallelism can be improved by increasing the number of cells. This approach to scalability has been explored in the Hurricane and Tornado operating systems developed at the University of Toronto [13]. Conclusion The experiments we have performed to date show that the fault containment mechanisms of Hive are effective and add low performance overheads when implemented in a commercial operating system. However, it is difficult to predict the reliability of a complex system before it is used ....

Unrau, R., Krieger, O., Gamsa, B., and Stumm, M. Hierarchical clustering: a structure for scalable multiprocessor operating system design. Journal of Supercomputing 9, 1/2 (Mar. 1995), 105--134.

No context found.

R. C. Unrau, O. Krieger, B. Gamsa, and M. Stumm, "Hierarchical clustering: A structure for scalable multiprocessor operating system design," The Journal of Supercomputing, vol. 9, no. 1--2, pp. 105--134, 1995. [Online]. Available: citeseer.nj.nec.com/unrau93hierarchical.html

No context found.

R. C. Unrau, O. Krieger, B. Gamsa, and M. Stumm, "Hierarchical clustering: A structure for scalable multiprocessor operating system design," The Journal of Supercomputing, vol. 9, no. 1--2, pp. 105--134, 1995. [Online]. Available: citeseer.nj.nec.com/unrau93hierarchical.html

....provide a framework for designing and implementing SMP software which is both highly concurrent and supports replication and partitioning of data so as to maximize locality. 1.2. 1 Hive and Hurricane The Stanford Hive operating system[7] and the University of Toronto s Hurricane operating system[37] were designed to address the locality issues in large scale SMP operating systems. Hive focused on locality, firstly as a means of providing fault containment and secondly as a means for improving scalability. Hurricane focused on the scalability and performance aspects of increased locality. ....

Ronald C. Unrau, Orran Krieger, Benjamin Gamsa, and Michael Stumm. Hierarchical clustering: A structure for scalable multiprocessor operating system design. The Journal of Supercomputing, 9(1-2):105--134, ???? 1995.

.... In core fault: 200 zsec Out of core fault: 800 zsec Base prefetch: 60 zsec per out of core page: 200 zsec per in core page: 30 zsec per in page table page: 10 zsec File System Operation Overhead Prefetch (per page) 70 sec Read Write (per page) 70 zsec cane operating system [33] running on the Hector shared memory multiprocessor [34] Hurricane is a hicrarchically clustered, micro kernel based operating system that is mostly POSIX compliant. It was largely irrelevant that the system was a multiprocessor; we chose this platform because the system has multiple disks ....

....instrumentation to enable us to produce the detailed statistics shown in subsequent sections. The basic characteristics of our experimental platform (with the instrumentation disabled) are shown in Table 1, and more detailed descriptions of the platform can be found in earlier publications [16, 33, 34]. We believe that our experimental results are conserwative for the following reasons: i) instru mentation is enabled for all the experiments, and Table 2: Description of applications. Memory Required Original of Execution Name Description Input Data Set Absolute Available Time (mins) BUK ....

R.. C. Unrau, O. Krieger, B. Gamsa, and M. Stumm. Hierarchical Clustering: A Structure for Scalable Multiprocessor Operating System De- sign. Journal of Supercomputing, 9(1/2):105 134, 1995.

.... 70 sec In core fault: 200 sec Out of core fault: 800 sec Base prefetch: 60 sec per out of core page: 200 sec per in core page: 30 sec per in page table page: 10 sec File System Operation Overhead Prefetch (per page) 70 sec Read Write (per page) 70 sec cane operating system [33] running on the Hector shared memory multiprocessor [34] Hurricane is a hierarchically clustered, micro kernel based operating system that is mostly POSIX compliant. It was largely irrelevant that the system was a multiprocessor; we chose this platform because the system has multiple disks ....

....instrumentation to enable us to produce the detailed statistics shown in subsequent sections. The basic characteristics of our experimental platform (with the instrumentation disabled) are shown in Table 1, and more detailed descriptions of the platform can be found in earlier publications [16, 33, 34]. We believe that our experimental results are conservative for the following reasons: i) instrumentation is enabled for all the experiments, and 7 Table 2: Description of applications. Memory Required Original of Execution Name Description Input Data Set Absolute Available Time (mins) ....

R. C. Unrau, O. Krieger, B. Gamsa, and M. Stumm. Hierarchical Clustering: A Structure for Scalable Multiprocessor Operating System Design. Journal of Supercomputing, 9(1/2):105--134, 1995.

.... operating system structure were mostly either done in the context of earlier non cachecoherent NUMA systems [8] or, as in the case of Plan 9, were not published [25] Two projects that were aimed explicitly at large scale multiprocessors were Hive [7] and the precursor to Tornado, Hurricane [30] . Both independently chose a clustered approach by connecting multiple small scale systems to form either, in the case of Hive, a more fault tolerant system, or, in the case of Hurricane, a more scalable system. However, both groups ran into complexity problems with this approach and both have ....

....available) and having the calling process migrate to that processor to execute the remote procedure. This approach would be prohibitive in today s systems with high cost cache misses and invalidations. 9 Concluding Remarks Tornado was built on our experience with the Hurricane operating system [30] . Hurricane employed a course grained approach to scalability, where a single large scale SMMP was partitioned into clusters of a fixed number of processors. Each cluster ran a separate instance of a small scale SMMP operating system, cooperatively providing a single system image. This approach ....

R. Unrau, O. Krieger, B. Gamsa, and M. Stumm. Hierarchical clustering: A structure for scalable multiprocessor operating system design. Journal of Supercomputing, 9(1/2):105--134, 1995.

....of a larger effort to investigate operating system design for large scale multiprocessors. Our operating system structure is called Hierarchical Clustering. The goal of Hierarchical Clustering is to support large scale applications without penalizing the performance of small scale applications [27]. The basic unit of structuring within Hierarchical Clustering is the cluster, which provides the full functionality of a tightly coupled small scale symmetric multiprocessor operating system. On a large system, multiple clusters are instantiated such that each cluster manages a unique group of ....

Ron Unrau, Michael Stumm, and Orran Krieger. Hierarchical Clustering: A Structure for Scalable Multiprocessor Operating System Design. Technical report, Computer Systems Research Institute, University of Toronto, 93.

....we intend to achieve it in the scalable, shared memory operating system we are currently designing and implementing, called Tornado. The design of Tornado is mostly driven from our experiences over the last 3 4 years in designing and implementing a first generation such system, called Hurricane [14], for our own locally developed multiprocessor, Hector [15] On this system, among other services, we developed a flexible parallel file system [7] that successfully exploits the techniques for flexibility we expect to apply in Tornado. We first describe the general techniques we use for ....

R. Unrau, M. Stumm, O. Krieger, and B. Gamsa. Hierarchical clustering: A structure for scalable multiprocessor operating system design. Journal of Supercomputing. To appear. Also available as technical report CSRI-268 from ftp.csri.toronto.edu.

....composition of resources created on its behalf. Also, the composition is dynamic and 2 On a large system, system data structures should be de clustered; that is, distributed (e.g. replicated, migrated, partitioned) among the different memory modules according to the demands placed on the data [13, 16]. can, in principle, be changed repeatedly by an application (assuming interface requirements are respected) Customizability In our building block framework, customizability can be achieved in a number of ways. First, given a particular composition, it is possible to exchange one building ....

R. Unrau, O. Krieger, B. Gamsa, and M. Stumm. Hierarchical clustering: A structure for scalable multiprocessor operating system design. Journal of Supercomputing, 9(1/2):105--134, 1995.

....are available. While the following development is applicable to demand driven systems in general, we are most familiar with parallel operating systems, and will therefore draw our examples from our experience in implementing an operating system for a scalable shared memory multiprocessor [10]. 2 Conditions for Scalability In this section, we develop a set of conditions for the scalability of demand driven systems. The development is based on an analysis of two fundamental performance metrics of computer systems: throughput and utilization. Section 2.1 reviews the definition of these ....

....factors. Consequently, in mitigating the eventual saturation of the resource as p is increased the system designer may wish to facilitate a limited version of replication that balances time against frequency of access. This is the approach used in the design of the Hurricane operating system [10]. 5 Evaluating Scalability In principle, the properties that form our scalability criteria are observable quantities. However, their measurement in real systems can be non trivial. For example, to prove a system is scalable, one must prove that s ck and V c are asymptotically bounded. For real ....

Ron Unrau, Orran Krieger, Benjamin Gamsa, and Michael Stumm. "Hierarchical Clustering: A Structure for Scalable Multiprocessor Operating System Design". Journal of Supercomputing, To appear 1995. This article was processed using the L a T E X macro package with LLNCS style

No context found.

R. Unrau, O. Krieger, B. Gamsa, M. Stumm. Hierarchical Clustering: A Structure for Scalable Multiprocessor Operating System Design. Journal of Supercomputing, 1995.

No context found.

R. Unrau, M. Stumm, O. Krieger. Hierarchical Clustering: A Structure for Scalable Multiprocessor Operating System Design. Technical Report CSRI-268, Computer Systems Research Institute, University of Toronto, March 1992.

No context found.

Ron Unrau, Michael Stumm, Orran Krieger, and Ben Gamsa. Hierarchical clustering: A structure for scalable multiprocessor operating system design. Journal of Supercomputing, 1995.

No context found.

Ronald C. Unrau, Orran Krieger, and Michael Stumm. Hierarchical clustering: A structure for scalable multiprocessor operating system design. Journal of Supercomputing, pages 105--134, March 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