DOUGLIS,F.,AND OUSTERHOUT, J. Process migration in the Sprite operating system. In Proceedings of the 7th International Conference on Distributed Computing Systems (ICDCS) (September 1987).  Home/Search   Document Not in Database   Summary   Related Articles   Check

....overall system processing power. Processes can then be distributed transparently among computers in the cluster. Clusters generally employ load sharing and process migration. Balancing load across processors or more generally across network nodes can be achieved via process migration mechanisms [3], where the state of a process is transferred between nodes. Transferring process state requires significant platform infrastructure support to handle platform differences between nodes. It may also limit applicability to programming languages based on virtual machines, such as Java. # ....

F. Douglis and J. Ousterhout, "Process Migration in the Sprite Operating System," in Proceedings of the 7 th International Conference on Distributed Computing Systems,(Berlin,WestGermany), pp. 18--25, IEEE, Sept. 1987.

....high amount of computation performed for each byte of information communicated between tasks[93] 2. 2 The Need for Process State Capture and Recovery Whereas the ability to capture and restore the state of a process has been available for some time in a variety of homogeneous distributed systems[4,20,54,64,70,86], this feature has been absent in most metacomputing environments. This is due primarily to the difficulty of implementing such a feature, not due to a lack of need for such a mechanism. 2.2.1 Load Balancing and Load Sharing It has long been recognized that adaptive load sharing as enabled by a ....

....of mechanisms and policies[62,64,76] Most homogeneous state capture mechanisms are implemented inside operating systems at the kernel level due to efficiency concerns and because a process s external state is more readily available at that level. For example, systems such as Charlotte[4] Sprite[20], DEMOS MP[70] and the V System[86] utilize kernel level state capture and recovery mechanisms to support process migration. Although these and other kernel level homogeneous state capture mechanisms differ in certain performance related respects, they share a common basic approach to capturing ....

F. Douglis and J. Osterhout, "Process Migration in the Sprite Operating System," in Proceedings of the 7th International Conference on Distributed Computing, pp. 18-25, 1987.

....mechanisms. 1 Introduction NUMA architecture [10, 1] is widely accepted as basic architecture for very highperformance computers because huge systems can be built by simply connecting many pairs of processing elements and local memories. Current parallelizing methods for on NUMA systems [11, 2, 8] and optimization techniques for parallel applications[12, 7] optimize execution of a single parallel application on a fixed system configuration. Therefore, running multiple parallel applications with competing resource allocation in general purpose environment is much less efficient than that is ....

F. Douglis and J. K. Ousterhout. Process Migration in Sprite Operating System. Proc. of the 7th Inter. Conf. on Distributed Computer Systems, September 1987.

....of Pages : 24 iv Chapter 1 Introduction NUMA[30, 7] systems are characteristic for that they may achieve very high performance by making huge systems by simply connecting many processing elements. There have been studies on supporting mechanisms[33, 8, 25] and optimization techniques for parallel applications[37, 24] These techniques optimize execution of a single parallel application. However, in general purpose environment, running every multiple parallel process with every resource allocation requirement satisfied is impossible. Running a ....

F. Douglis and J. K. Ousterhout. Process Migration in Sprite Operating System. Proc. of the 7th Inter. Conf. on Distributed Computer Systems, September 1987.

....the migration. The need for process migration was discussed in the late 1970s by Solomon and Fenkel (1979) Successful implementations have been presented in (Lazowska et al. 1981; Rashid and Robertson, 1981; Powell and Miller, 1983; Walker et al. 1983; Theimer et al. 1985; Artsy et al. 1986; Douglis and Ousterhout, 1987, 1989; Zayas, 1987; Walker and Mathews, 1989) Process migration is considered a very difficult task to construct. Problems that have to be handled Include (Goscinsky, 1991) 1. determine the process state, detaching a process from its current environment, transferring it with all its relevant ....

Douglis, F., Ousterhout, J. (1987). Process migration in the Sprite operating system. In: Proceedings of the 7th International Conference on Distributed Computing Systems, Berlin, west Germany, pp. 18-25.

....and a preliminary version is running on SUN 2 and SUN 3 workstations. Contact: Fred Douglis or Michael N. Nelson, Computer Science Division, University of California at Berkeley, 571 Evans Hall, Berkeley, CA 94720. For principal investigation, contact John Ousterhout at UC Berkeley. References: [308], 309] 310] 311] 312] 313] 314] 2.49 SWALLOW Main Goal SWALLOW is a distributed file system supporting highly reliable object oriented data storage. It provides all kind of transparency levels, i.e. those of location, access, concurrency, failure, and replication. Advantages ....

F. Douglis and J. Ousterhout, "Process Migration in the Sprite Operating System", In Proc. 7th Int. Conf. on Distributed Computing Systems, pages 18--25, Berlin, West Germany, September 1987.

....the mobile host uses end to end TCP when it is connected to its home network. The same TCP connection is however split at a mobility support gateway (MSG) when the MH moves away. 8. 3 Handoffs and State Migration Although the idea of migrating objects and processes has been around for a while [27, 50], we are not aware of any work preceding ours that allows migration of active transport and higher layer connections from one machine to another. Handoff models in cellular systems and mobile IP schemes [41, 72, 73, 74, 87, 91] typically deal with lower layer handoffs which do not involve ....

F. Douglis and J. Ousterhout. Process migration in the Sprite operating system. In Proc. of the 7th Intl. Conf. on Distributed Computing Systems, pages 18--25, 1987. 169

....is optimal if all job sizes divide each other, e.g. if they are powers of two [5] It is debatable whether this algorithm is realistic, because of the expected overhead, especially on distributed memory machines. It is true that systems that support migration have been implemented successfully [1, 6], but these systems do not attempt to perform migration at such a high rate. However, this algorithm is useful as a bound on the performance that is obtainable. 3 Workload Modeling The most straightforward way to evaluate scheduling algorithms without a full scale implementation is through ....

F. Douglis and J. Ousterhout, "Process migration in the Sprite operating system". In 7th Intl. Conf. Distributed Comput. Syst., pp. 18--25, Sep 1987.

....for optimistic systems, this distinction has not arisen before. Some form of object migration is needed if existing objects are to be shifted among the system s nodes. Many existing distributed systems support process migration, including Accent (Zaya 1987) Locus (Popek and Walker 1985) Sprite (Douglas and Ousterhout 1987), DEMOS MP (Miller 1987) the V kernel (Theimer et al. 1985) KLOX (Li 1988) and the Rand Time Warp implementation (Burdorf and Marti 1990) Of these, only KLOX and Rand Time Warp have access to virtual time in their migration mechanism, and neither use the concept of temporal splitting to assist ....

F. Douglas and J. Ousterhout. 1987. "Process Migration in the Sprite Operating System." In 7th International Conference on Distributed Computing Systems.

....arbitrary migration policies. This paper describes the implementation and first experiences. 1 Introduction Migration of objects in distributed systems is found in quite a few research operating systems. Prominent special cases for object migration are file caching and process migration [Douglis87, Barak85]. In most cases, migration is used to enhance the performance of systems by balancing the load among a number of computing nodes or by reducing communication overhead. Specialized to support load balancing, most migration mechanisms possess severe limitations: ffl migration is not available for ....

F. Douglis and J. Ousterhout, Process Migration in the Sprite Operating System, Seventh Int. Conf. on Distributed Computing Systems, Sep 1987, pages 18-25

....the plan, as well as a good understanding of the applications, perhaps only a performance robust scan read and sort write would be needed, as the mechanisms therein would naturally cope with any performance variations to the system. 13.3. 5 Migration Though ignored in this work, process migration [13, 47, 83, 97, 98, 120] could be utilized within our environment to provide some amount of performance availability. In the current system, process placement is presumed to be static; robustness is added to applications by facilitating dynamic movement of data to and from faster entities. However, a system that also ....

....unreliable in terms of performance, any remaining work could be moved to other, better responding nodes. The difficulty with this strategy could be in getting the data off of the slow node. Transparent process migration traditionally has been difficult to implement, for reasons enumerated in [47]. This may be why no current operating systems deliver a migration package as a part of their standard system. However, in a specific programming environment such as River, non transparent migration may provide an easier path towards an implementation, where the user is required to program extra ....

Fred Douglis and John K. Ousterhout. Process Migration in the Sprite Operating System. In Proceedings of the 7th International Conference on Distributed Computing Systems, pages 18--25, Berlin, West Germany, September 1987. IEEE.

....(2) an application typically internalizes several operating system handles (such as socket identifiers and IP addresses) which stop being relevant upon migration. Existing approaches deal with the above problems either by relying on extensive modifications to OS structures to support migration [6, 12, 13, 10], or by requiring the use of a new application programming interface (API) 2, 4, 3, 7] whose implementation isolates the application from the consequences of migration. Neither of these choices is ideal because they cannot be applied to existing OSes and applications. Moreover, most such ....

....multiprocessor system. 2.1 Related Work Previous approaches for maintaining network connectivity fall into two broad categories: modifications to the OS network protocol stack, and introduction of a new API for accessing underlying resources. Modifying the OS network layer. Several researchers [13, 12, 6] have successfully demonstrated transparent network connectivity by across process migrations by incorporating changes to kernel data structures and protocols. For instance [13, 12] the implementation of network migration on top of the Chorus operating system modifies the network manager so that ....

Douglis, F. and Ousterhout, J. Process migration in the sprite operating system. In Proc. of 7th Intl. Conf. on Distributed Computing Systems, pages 18--25, 1987.

....process migration performance for RHODOS and the described systems. Finally, Section 7 concludes this report, and indicates the direction of our future work. 2 Design Issues Our initial research [Goscinski 91] and an analysis of several process migration systems created [Artsy and Finkel. 89] Douglis and Ousterhout 87] Douglis and Ousterhout 91] Jul et al. 88] Powell and Miller 83] Smith 88] Zayas 87] show that all use differing designs. In order to achieve the goal specified in the introduction we propose that the following issues should be considered when designing a process migration facility. ....

....(a RHODOS process has two by default) i. All hardware is connected by a 10 Mbit Ethernet LAN. ii. Based on paging time of 15.24 ms [Milojicic 94a] pp76. iii. more recent results are published, however on different network hardware to the other systems. iv. Based on paging time of 15 ms [Douglis and Ousterhout 87] Table 6.1 Comparison of migration for different systems Distributed Operating System how address space is copied hardware i time to migrate (in milliseconds) freeze time for 100K process freeze time for worst case (100K) Amoeba direct copy 386 PC N A 1.5 sec 1.5 sec Chorus ....

[Article contains additional citation context not shown here]

F. Douglis, J. Ousterhout. Process Migration in the Sprite Operating System. Proceedings of the 7 th international Conference on Distributed Computing Systems, Berlin, September.

....be stopped on the original host and a new copy re instantiated on the new host. There is a limited mechanism for a final state transfer by a single buffer transfer from the old instance to the new instance. This approach avoids many of the details involved in checkpointing and process migration [11], while retaining most of the benefits. Filters need to be structured appropriately to handle such events. For cases when this is not desirable, a filter can be pinned to a particular host, which means the filter will always be placed on that host. This host affinity is useful for some situations, ....

F. Douglis and J. Ousterhout. Process migration in the Sprite operating system. In Proceedings of the Seventh International Conference on Distributed Computing Systems, pages

....application. This paper describes such a reconfiguration mechanism in the D CVM (Dynamic Coherent Virtual Machine) 1] distributed shared memory (DSM) system. D CVM implements reconfiguration through thread migration. Thread and process migration has long been used as a loadbalancing mechanism [2, 3] in parallel and distributed systems. However, DSMs usually have much higher communication requirements than message passing systems, implying that good thread migration policies in this domain must also account for communication behavior. Consider a page based software DSM. If threads on ....

F. Douglis and J. Ousterhout, "Process Migration in the Sprite Operating System," in Proceedings of the 7th International Conference on Distributed Computing Systems, September 1987.

....Rover, on the other hand, is designed to be more flexible. Depending on the power of the mobile host and available bandwidth, Rover dynamically adapts and moves functionality between the client and the server. RDOs can be viewed as simple agents [27] or as a light weight form of process migration [28], 29] 30] 31] Other forms of code shipping include Display PostScript [32] Safe Tcl [33] Active Pages [34] Dynamic Documents [35] and LISP Hypermedia [36] RDOs are probably closest to Telescript [37] Ousterhout s Tcl agents [38] and Java [39] Most differences between RDOs and these ....

F. Douglis and J. Ousterhout, "Process migration in the Sprite operating system," in Proc. of the 7th International Conference on Distributed Computing Systems, Berlin, West Germany, Sept. 1987, IEEE, pp. 18--25.

....of standard services for creating, locating, registering, and storing an object 2 . 3.2 Distributed Operating Systems The operating system approach addresses a dioeerent aspect of distribution by migrating processes for load balancing, fault tolerance, and resilience. Systems such as Sprite[DO87] V[Che88] and Locus[PE86] provide built in operating system support to freeze the run time computation of a process, migrate the process to a remote site, and unfreeze the process. Each migration entails leaving a proxy behind to forward I O and communications to the new site. By providing ....

Fred Douglis and John Ousterhout. Process migration in the sprite operating system. In Proc. of the 7th Intl. Conf. Distributed Computer Systems, pages 1825, 1987.

....a minimal freeze time. But, since the virtual memory of the process is not transferred, the process will take longer to execute because of the page faults. Another disadvantage is that this scheme leaves residual dependencies on the source machine. The enhanced lazy copying scheme, used by Sprite [33], overcomes the disadvantage of lazycopying by flushing the dirty pages to a file server (on the swap space of the process) thus, no residual dependencies are left on the source machine after migration. On the other hand, a page fault may be more costly than with the previous approach. 3.3 Load ....

....workstations connected by an 80 Mbit sec token ring based LAN. A 100 kilobytes process requires 55 187 milliseconds to migrate from one processor to another [25] In Sprite, running on a Sun 3 workstation, 600 milliseconds are required to migrate a 100kilobyte process with three open files [33]. Unfortunately, there are no published specific measurements of Demos MP and Locus which we could list. 5 Concluding Remarks The design issues of process migration mechanisms in distributed systems have been presented. Analysis of these systems shows that process migration is expensive and not ....

Fred Douglis and John K. Ousterhout. Process Migration in the Sprite Operating System. In Proceedings of the 7th International Conference on Distributed Computer Systems, pages 18--25, September 1987.

....after adaptation all affect the cost of adaptation. 6 Related work Various systems have been developed to allow sequential computations to use idle time on networked nodes. These systems include, among others, Butler [20] Condor [17] and many process migration systems such as, e.g. Sprite [8]. Our work distinguishes itself from these systems by its support for parallel computations. Cilk NOW [5] Dataparallel C [19] Piranha [6] and various migration based systems (e.g. Millipede [12] or versions of PVM [15] support adaptive parallel computation on NOWs. Blumofe and Lisiecki [5] ....

F. Douglis and J. Ousterhout. Process Migration in the Sprite Operating System. In Proceedings of the 5th International Conference on Distributed Computing Systems, pages 18--25, September 1987.

....algorithm differs in that it does not go as far as balancing the system but just make sure that there is no idle host while a job is queueing in another host. The DGP algorithm is based on non preemptive scheduling. Preemptive scheduling in distributed systems, known also as process migration [9] involves a considerable amount of computation when transferring a process and all its environment (such as virtual address space, cached writable files) from one host to another. Furthermore, it is difficult to implement on Unix systems, and it is even more complex for jobs that generate several ....

F Douglis and J Ousterhout. Process migration in the sprite operating system. Proceedings of the Seventh International Conference on Distributed Computing Systems, pages 18 -- 25, September 1987.

....Accent migrates a 100kB process in approximately 13,000ms. The Accent process migration time formula in milliseconds is: T otalT ime = 1180 115 MemorySizeKiloBytes (3:6) 3.3.22 Sprite Sprite [OCD 88] is a network operating system with a distributed file system. Sprite process migration [DO87] Dou89] DO89] Dou90] DO91] takes advantage of the Sprite network file server to gain much of the benefits of demand paged memory transfer while simultaneously eliminating the long term memory residual dependencies. The integration of file server support into process migration is another milestone ....

....memory pages and dirty file cache blocks to the file server. The new host demand pages the file blocks and memory pages from the file server. Since Sprite applications are always dependent on the Sprite file server, this approach does not add a new machine dependency. Sprite process migration [DO87] Dou89] DO89] Dou90] DO91] swaps the cost of the file server flush for the elimination of the address space residual dependency. Sprite does not eliminate all residual dependencies on the old host machine, because the system forwards some system calls to the home machine for execution, such as ....

Douglis and Ousterhout. Process migration in the Sprite operating system. In 7th International Conference on Distributed Computing, pages 18-- 25, 1987.

No context found.

F. Douglis and J. Ousterhout, `Process migration in the Sprite operating system', Proceedings of the 7th International Conference on Distributed Computing Systems, IEEE, Berlin, West Germany, September 1987, pp. 18--25.

No context found.

DOUGLIS,F.,AND OUSTERHOUT, J. Process migration in the Sprite operating system. In Proceedings of the 7th International Conference on Distributed Computing Systems (ICDCS) (September 1987).

No context found.

F. Douglis and J. Ousterhout, "Process Migration in the Sprite Operating System," in Proceedings of the 7 th International Conference on Distributed Computing Systems, September 1987.

No context found.

F.Douglis and J. Ousterhout, "Process migration in the Sprite Operating System," Proc. of the 7th Int'l Conf.onDistributed Computing Systems,pp. 18-25 IEEE Computer Press, (September 1987).

First 50 documents  Next 50

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