Mor Harchol-Balter and Allen B. Downey. Exploiting process lifetime distributions for dynamic load balancing. In Proceedings of the ACM Sigmetrics Conference on Measurement and Modeling of Computer Systems, pages 13--24, May 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....up a link in the core (2.5 Gbit s and above) on its own. Second, it has been shown that whereas flow session arrivals are Poisson (or close to Poisson) 78, 45] flow sizes are not exponential, but rather heavy tailed, and thus they are closer to a Pareto distribution than an exponential one [84, 57, 183]. This chapter evaluates the end user response time with consideration of the characteristics of the current Internet. 3.3 LANs and shared access networks I will start with some examples to illustrate what may happen when circuit or packet switching is used. I will use a simple example to ....

....size variances, especially for high loads when the job queue is often non empty and the reordering of the queue makes a difference. However, CS SJF cannot avoid the eventual hogging by long flows when the job size variance is high, and then it performs worse than PS PrSh. It has been reported [84, 57, 183] that the distribution of flow durations is heavy tailed and that it has a very high variance, with an equivalent a greater than 0.999. This suggests that PS PrSh is significantly better than either of the circuit switching disciplines. I have further verified this conclusion using a ....

Mor Harchol-Balter and Allen B. Downey. Exploiting process lifetime distributions for dynamic load balancing. ACM Transactions on Computer Systems, 15(3):253-- 285, 1997.

....supply prediction i.e. load prediction is found in some form or the other in every adaptive systems. Typically, prediction is implicit in these systems: offered load or other system indicators are used as control signals that drive load balancing [58] scheduling [27] process migration [28, 45], or resource allocation [89] A few systems use explicit resource prediction. The Running Time Advisor [25] predicts the running time of a job on any host in a distributed system, given its nominal execution time: i.e. it predicts resource supply and performance, but the application must ....

....example, autoregressive linear models [10] take the variance as well as the mean of recent observations into account. Such techniques could improve the accuracy of my relatively simple supply predictors (Section 5.5) Resource demand prediction is comparatively rare. Some load balancing systems [58, 45] estimate the CPU demand (execution time) of a job from an analytically or empirically derived probability distribution. However, they do not make use of any dynamic, instancespecific information such as the job s runtime parameters: at most, they use the job s current lifetime to predict its ....

Mor Harchol-Balter and Allen B. Downey. Exploiting process lifetime distributions for dynamic load balancing. In Proceedings of the Joint International Conference on Measurement and Modeling of Computer Systems (ACM SIGMETRICS '94), pages 13--24, Nashville, TN, May 1994.

....policies in distributed memory systems. By comparing the performance of share scheduling policies on multiprocessors, for which we assume that job migration is free, and distributed systems, one can gain insight into the controversial benefits of job migration in distributed systems [36, 89]. A final restriction in this thesis is that we do not consider jobs consisting of multiple tasks, such as for instance studied in [2] Such jobs are often used to model parallel programs. Although in this thesis we restrict ourselves to the application of share scheduling in distributed systems, ....

M. Harchol-Balter and A.B. Downey. Exploiting Process Lifetime Distributions for Dynamic Load Balancing. ACM Transactions on Computer Systems, 15(3):252--285, 1997.

....tails as there are no linear regions that spans several orders of magnitude. Note that the limited and seemingly non dispersive distributions we see are in stark contrast to data from interactive workstations, where the tail of the distribution is indeed heavy and follows a Pareto distribution [17, 12]. There is no reason to expect such long jobs to be absent on supercomputers on the contrary, sans administrative restrictions, long jobs may be expected to dominate the workload. Finally, we note that even if the distribution of job runtimes does not conform to the formal definition of ....

....of dispersive distributions and the non negligible probability of sampling very high values have significant implications on systems. For example, Harchol Balter and Downey have shown that when the distribution of job runtimes has a heavy tail, migration for load balancing can be very beneficial [12]. This is based on the fact that a small number of jobs use more CPU time than all the others; the benefits come from identifying these jobs and migrating only them. This contradicts evaluations based on a more moderate distribution, in which migration did not lead to significant benefits [7] In ....

M. Harchol-Balter and A. B. Downey, "Exploiting process lifetime distributions for dynamic load balancing ". ACM Trans. Comput. Syst. 15(3), pp. 253--285, Aug 1997.

....corresponds to a slower descent and hence produces a heavier tail. When Leland and Ott measured UNIX task lifetimes in 1985, they found that ff ranges from 1:05 to 1:25. More recently, Harchol Balter and Downey performed an independent study of process lifetimes finding ff to range from 0:8 to 1:3 [17]. We will focus on ff s in this range to be consistent with this and other work [10, 11, 16] 15 We believe research using heavy tailed distributions is important both because of the stated evidence that they abound in queuing systems and because tasks with heavy tailed service times have a few ....

M. Harchol-Balter and A. Downey. Exploiting Process Lifetime Distributions for Dynamic Load Balancing. In ACM Transactions on Computer Systems, 15(3):253-285, August 1997.

....Science Foundation under grant CCR 0098278. 1 Back end Nodes . Arriving tasks Dispatcher Front end Figure 1: Model of a clustered web server. heavy tail distributions [2, 3, 4, 5] As a consequence, load balancing strategies that perform well under light tail workloads become inadequate [15]. In this paper, we propose a new, easy to implement scheduling policy for a clustered web server, whose abstraction is illustrated in Figure 1. With our new scheduling policy, called EquiLoad, the size of each incoming request determines the identity of the back end server that will eventually ....

....such as minimizing the expected response time. Identifying the best server to which assign each request requires detailed information about the distribution of the service demands. In the past two decades, there has been a signi cant research e ort in task scheduling and load balancing (see [15] and references therein) The basic assumption in much of this work is that the service demands of the various tasks are governed by an exponential distribution. In contrast to the above assumption, however, there is very strong evidence that the size of a web document, and accordingly its service ....

[Article contains additional citation context not shown here]

M. Harchol-Balter and A. Downey. Exploiting process lifetime distributions for dynamic load balancing. ACM Transactions on Computer Systems, 15(3):253-285, Aug. 1997.

....needs more dynamic system information including the status of buffer caches. Quantifying lifetimes of jobs without detailed breakdowns: The lifetime of a job has been used as an important factor in load sharing designs. Which process to be migrated in existing scheduling policies (see e.g. [6]) is considered by predicting the lifetime of CPU intensive jobs. The lifetime of a job mainly consists of CPU time, memory paging time, and I O service time. Detailed breakdowns of the lifetime into these major portions will provide a multi dimensional set of information for load sharing ....

.... to migrate the job with the highest page fault rate (simplified as pf rate in figures) working set size based: to migrate the job with the largest working set size (simplified as worksize in figures) age based: to migrate the job with the oldest age with a migration cost estimation [6] (simplified as age in figures) CPU memory I O based: to migrate the job with the highest activities of CPU, memory, and I O calculated in (5) simplified as combined in figures) In addition, we have also compared the performance of the above schemes with that without using load sharing, ....

[Article contains additional citation context not shown here]

M. Harchol-Balter and A. B. Downey, "Exploiting process lifetime distributions for dynamic load balancing", ACM Transactions on Computer Systems, Vol. 15, No. 3, 1997, pp. 253-285.

.... reported for several previous parallel workloads [FeNi95, WLF 96, Gib97b] Plots of the complement of the runtime distribution on a log log scale (not shown) reveal that the observed distributions of runtime are not heavy tailed, in contrast to the process runtimes observed on Unix systems [HoDo97] and a hypothetical model of parallel job 0 0.2 0.4 0.6 0.8 1 10 0 10 1 10 2 10 3 10 4 10 5 30s 1m 5m 10m 1h 5h 10h 50h 100h 400h cumulative fraction of jobs actual runtime (minutes) Jan00 (5752) Feb00 (6130) Mar00 (7211) Oct00 (5628) Nov00 (4870) Dec00 (4182) a) ....

M. Harchol-Balter and A. B. Downey. "Exploiting Process Lifetime Distributions for Dynamic Load Balancing". ACM Trans. on Computer Systems, 15(3):253--285, August 1997.

....agents can be useful relates to load balancing. The aim of a load balancing strategy is to adapt to the utilisation and performance of the machines available and to the requirements of the agents. In operating systems, researchers investigated how load balancing can be optimised with mobility [11, 25]. In this section we survey how load balancing with and without mobility is used for agent systems. We can distinguish systems according to two criteria: static vs mobile, i.e. are tasks assigned to hosts once and then stay there or can they migrate if it turns out at a later stage that it is ....

....to a static placement scheme is achieved. 2.2 Model based Load balancing Model based approaches to load balancing are much rarer, as it involves two very challenging problems: How to obtain an initial model and how to adapt the model as time passes. In the area of operating systems, some authors [11, 25] explored this approach and use distributions of CPU load and expected process lifetime to decide if and when to migrate. A similar idea is used in Challenger [12] It implements load balancing with a market approach. However instead of using money agents make bids giving the estimated time to ....

[Article contains additional citation context not shown here]

M. Harchol-Balter and A.B. Downey. Exploiting process lifetime distributions for dynamic load-balancing. ACM Transactions on Computer Systems, 15(3):25385, 1997.

....earlier than expected. Predicting when running jobs complete requires an estimate of the remaining run time of each job. One rule of thumb for estimating job run times based on a study of job lifetime distributions is that the median remaining lifetime of a process is equal to its current age [27, 33]. Enforcing run time limits on jobs can limit the penalty for mispredicting run times. Users submit their jobs to queues with configured run time limits, choosing the appropriate queue based on their own estimates of their jobs behavior. The scheduler can then use these run time limits as ....

....to I O, checkpoint size, etc. users can choose simple metrics to request allocations more suited to the needs of their jobs. 7.1. 4 Migrating to Improve Goodput The benefits of job migration and the policies under which jobs should be migrated have been extensively studied (see, for example, [16, 17, 27, 32, 33]) We briefly discuss potential applications of the goodput framework to the job migration question here. The benefit of migrating to a site that promises better performance must be compared to the expected cost of migration. From the goodput perspective, the cost is the time the job spends ....

M. Harchol-Balter and A. Downey. Exploiting process lifetime distributions for dynamic load balancing. ACM Transactions on Computer Systems, 15(3), August 1997.

....vary only if the application machine load varies. The predic2. Multithreaded SFOs are the subject of future work. 10 tions of application computation time and communication are based on a weighted average of the recent past over a time window. More recent values appear to be a strong predictor [7][16] 17] and an exponential decay function is used to model this relationship [15] The SFO determines the application computation values, comp, by simply timing the interval between successive requests from the application. Similarly, the communication times, comm, are computed by timing requests ....

M. Harchol-Balter, A.B. Downey, "Exploiting Process Lifetime Distributions for Dynamic Load Balancing," SIGMETRICS, 1996.

....approximations, reflected Brownian motion, transient behavior 2 G. Hjalmtysson, W. Whitt Periodic Load Balancing 1. Introduction There is now a substantial literature on dynamic multiprocessor load balancing; e.g. see Eager, Lazowska and Zahorjan [14] Hajek [20] Harchol Balter and Downey [21], Leland and Ott [28] Willebeck LeMair and Reeves [49] Zhou [53] and references therein. The basic scheme is to move jobs from a highly loaded originating processor to another more lightly loaded processor. There can be significant overhead associated with this load balancing, but it is ....

.... originate, but migration of jobs in process is now beginning to be used as well, e.g. see Barak, Shai and Wheeler [8] There is typically substantially more overhead with migration of jobs in process, but it has been shown to yield significant performance improvement by Harchol Balter and Downey [21]. A di#culty with any form of dynamic load balancing, however, is that it involves real time control, requiring continuous maintenance of state information. It is thus natural to consider whether it is possible to achieve much of the load balancing benefit with less work. Hence, in this paper we ....

[Article contains additional citation context not shown here]

M. Harchol-Balter and A. B. Downey, Exploiting process lifetime distributions for dynamic load balancing. Proceedings SIGMETRICS `96, 1996.

....is an inherently slippery property how can it ever be verified But then, independence can never be proven in the physical world, either, and few have difficulty accepting its use in modeling. The evidence for heavy tails is widespread, including CPU time consumed by Unix processes [LO86, H BD96] sizes of Unix files [Irl93] compressed video frames [GW94] and World Wide Web items [CB97] and bursts of Ethernet [WTSW95] and FTP [PF95] activity. Note that the log normal distribution discussed in the previous item is not a heavy tailed distribution, yet these two invariants are not in ....

M. Harchol-Balter and A. Downey, "Exploiting process lifetime distributions for dynamic load balancing," Proc. SIGMETRICS '96, May 1996.

....Weather Service [28] or Topology d [24] provide measurements and predictions to help applications make scheduling decisions. This paper characterized one such measurement (the Unix load average) and studied how to predict it. While considerable effort has gone into characterizing workloads [12, 20, 16], the focus has been on load sharing systems [11] which schedule all of jobs in a distributed system, and not on application centric scheduling. An important assumption in load sharing and balancing systems is that current load is a predictor of future load. This work shows that that assumption ....

HARCHOL-BALTER, M., AND DOWNEY, A. B. Exploiting process lifetime distributions for dynamic load balancing. In Proceedings of ACM SIGMETRICS '96 (May 1996), pp. 13-- 24.

....packet header) 40 (one ACK) bytes sec = 120 bytes sec = 960 bits sec. A useful fact about processes is that they can be characterized as either CPU bound 2 If the time lag between successive data packets is more than 1 2 second, then there is one ACK for every data packet. 5 or I O bound [63, 40]. In process duplication or re computation, processes which consume little CPU time (I O bound processes) but produce lots of output are duplicated at the other peer. Whereas performing re computation on the server to generate the output les would be inecient in a LAN based environment, the ....

....of output per process. The traces which we used satis ed the second criterion perfectly, but the rst criterion was satis ed only approximately in the sense that arrival and departure times for a process were available but the CPU time was not. There were traces available containing CPU times [40], but since they did not contain any information about the outputs of processes, they were not of immediate use for our purposes. The traces were gathered on a Sparc station 10 on SunOS 4.1.3 [37] There were two traces available. The rst one (called Harp) is a result of two graduate students ....

M. Harchol-Balter and A. Downey. Exploiting process lifetime distributions for dynamic load balancing. In Proceedings of the ACM SIGMETRICS Conference on Measurement and Modeling of Computer Systems, May 1996.

....in determining which properties are relevant and whether all of the possibly relevant properties have been discovered. Both the networking and load balancing communities have histories of choosing the wrong properties, resulting in unfortunate system designs that have only recently been corrected [13, 8]. When we studied the statistical properties of host load signals, we found complex properties such as a strong autocorrelation structure, self similarity and epochal behavior [5] Predictive models use these properties to create forecasts, and thus it is vital to get them right in the model that ....

Harchol-Balter, M., and Downey, A. B. Exploiting process lifetime distributions for dynamic load balancing. In Proceedings of ACM SIGMETRICS '96 (May 1996), pp. 13--24.

....distributed with mean 5 sec. comparable to the CPU time requested by parallel processes) Experiments for other values have been performed as well, but are not discussed here because of space constraints. In order to evaluate the impact of heavy tailed distributions of process CPU times [5, 12], we have also performed experiments where the bounded Pareto distribution [10] has been used for the CPU demand, but we do not discuss them here because the results are not significantly different. The average completion time of parallel and sequential applications are shown in Figs. 2 to 5 ....

A. Downey and M.Harchol-Balter. Exploiting Process Lifetime Distributions for Dynamic Load Balancing. ACM Trans. on Computer Systems, 15(3):253--385, August 1997.

....by any server. Such a system can be considered as an abstraction of a distributed web server [7, 10, 19] Balancing the load across the back end servers is critical for performance [7] In the past two decades, there has been a signi cant research e ort in task scheduling and load balancing (see [11] and references therein) The This work has been supported by National Science Foundation under grands EIA 9974992 and EIA9977030, and by the National Aeronautics and Space Administration under NASA Grant NAG 1 2168. basic assumption in much of this work is that the service demands of the ....

M. Harchol-Balter, and A. Downey. Exploiting Process Lifetime Distributions for Dynamic Load Balancing. ACM Transactions on Computer Systems, 15(3), pp. 253-285, Aug. 1997.

....are compared with random and diffusion algorithms, and show that our model is efficient, adaptable and scalable with a low system overhead. 1 Introduction Dynamic process migration is the main load balancing mechanism to react to changes both in computation pattern and in system environment [10]. Migration of non communicating processes, by means of dynamic load balancing, based on computational needs is a well known and difficult problem [4] Taking into account usage of resources (such as communication between processes) make migration algorithms more intricate. They need to consider ....

Mor Harchol-Balter and Allen B. Downey. Exploiting process lifetime distributions for dynamic load balancing. Technical Report UCB CSD 95 021, Berkeley, University of California, May 1995.

....such tails as there is no linear regions that spans several orders of magnitude. Note that the limited and seemingly non dispersive distributions we see are in stark contrast to data from interactive workstations, where the tail of the distribution is indeed heavy and follows a Pareto distribution [21, 16]. There is no reason to expect such long jobs to be absent on supercomputers on the contrary, sans administrative restrictions, long jobs may be expected to dominate the workload. Finally, we note that even if the distribution of job runtimes does not conform to the formal de nition of having a ....

....of dispersive distributions and the non negligible probability of sampling very high values have signi cant implications on systems. For example, Harchol Balter and Downey have shown that when the distribution of job runtimes has a heavy tail, migration for load balancing can be very bene cial [16]. This is based on the fact that a small number of jobs use more CPU time than all the others; the bene ts come from identifying these jobs and migrating only them. This contradicts evaluations based on a more moderate distribution, in which migration did not lead to signi cant bene ts [8] In ....

M. Harchol-Balter and A. B. Downey, \Exploiting process lifetime distributions for dynamic load balancing". ACM Trans. Comput. Syst. 15(3), pp. 253-285, Aug 1997.

.... than migrating the address space of a running process (which could be partially paged out) We feel that most of the load balancing benefit of remote processes can be obtained by positioning the process at execution time (as do the Plan 9 researchers [19] although Harchol Balter and Downey [11] have found benefits from moving running processes) We plan to implement migration of running processes later, in particular to off load processes from a node scheduled for maintenance; at this time we will provide rfork and migration (which are very similar, since a rfork is basically a fork and ....

M. Harchol-Balter and A. B. Downey, "Exploiting Process Lifetime Distributions for Dynamic Load Balancing, " Proceedings of ACM Sigmetrics '96 Conference on Measurement and Modeling of Computer Systems, May 23-26 1996, pp 13-24.

....the job completes service. Service requirement is just the CPU requirement in our case this is the response time minus the queueing time. All means are per job averages. Mean slowdown is important because it is desirable that a job s response time be proportional to its processing requirement [8, 3, 13]. Users are 1 DISPATCHER HOST HOST HOST HOST 1 2 3 4 OUTSIDE ARRIVALS FCFS FCFS FCFS FCFS Figure 1: Illustration of a distributed server with 4 host machines, each of which is a multiprocessor. likely to anticipate short delays for short jobs, and are likely to tolerate long ....

Mor Harchol-Balterand Allen Downey. Exploiting process lifetime distributions for dynamicload balancing. ACM Transactions on Computer Systems, 15(3), 1997.

....all of our suggested policies exploit. Lastly, we evaluate the policies based on the probability that a deadline is met, rather than on the average time to complete all the work. Finally, our work is also related in spirit to earlier results on load balancing in distributed systems (e.g. [12, 25, 4, 6]) especially those that dealt with real time tasks (e.g. 9, 1] However, the former had as a goal minimizing response time, rather than meeting real time deadlines, while the latter addressed workloads in which each task had its own deadline, rather than our situation in which there is a ....

M. Harchol-Balter and A. B. Downey. Exploiting Process Lifetime Distributions for Dynamic Load Balancing. In Proceedings of the ACM SIGMETRICS Conference, pages 13--24, May 1996.

....can be exploited to reduce the overheads of certain control mechanisms. Observations of heavy tailed lifetime distributions of UNIX processes have been similarly exploited in the context of processor load balancing, where migrating long lived jobs can significantly reduce migration overhead [16]. Most notably in the networking context, variability in flow duration has been the basis of several techniques that reduce router forwarding overheads by establishing hardware switching paths for long lived flows [17 19] These schemes classify arriving packets into flows and apply a trigger ....

M. Harchol-Balter and A. Downey, "Exploiting process lifetime distributions for dynamic load balancing," ACM Transactions on Computer Systems, vol. 15, pp. 253--285, August 1997.

....is denoted as T migr . If we knew the non idle state would last long enough to make migration advantageous, an immedi 3 ate migration would be the best choice. But because we don t know when the non idle state will end, we have to predict it. We use the observations of Harchol Balter and Downey[5], and Leland and Ott[8] which states that the median remaining life of a process is equal to its current age. So if a process has run for T units of time, we predict its total running time will be 2T. Our use of this predictor is somewhat different since we use it to infer the duration of a ....

M. Harchol-Balter and A. B. Downey, "Exploiting process lifetime distributions for dynamic load balancing, " SIGMETRICS. May 23-26, 1996, Philadelphia, PA, pp. 13-24.

....show the necessity of mechanisms like co scheduling and migration of parallel jobs to idle workstations, but they do not predict contention effects. In fact, both [16] and [17] propose co scheduling strategies (or an approximation) for parallel applications executing in network of workstations. In [26], Harchol Balter and Downey show that preemptive migration, in which running processes may be suspended, moved to a remote host, and restarted, can minimize the effect imposed by contention on the hosts. Leutenegger and Sun [30] have created a model to show the viability of non dedicated ....

M. Harchol-Balter and A. B. Downey, "Exploiting Process Lifetime Distributions for Dynamic Load Balance", in Proceedings of ACM SIGMETRICS'96, pp. 13-24, May 1996.

....A synthetic compute intensive job was periodically invoked on every node. Processor load was generated using models developed by Harchol Balter and Downey, whose measurements indicate Poisson interarrival times, with job duration determined by a combination of exponential and Pareto distributions [12]. Because we are interested in environments which support compute and data intensive computations, higher parameters were used for the load generators than would be used to represent typical interactive systems. Assuming that our target environment is a cluster or group of workstations in a single ....

HARCHOL-BALTER, M., ANDDOWNEY, A. B. Exploiting process lifetime distributions for dynamic load balancing. In 1996 ACM SIGMETRICS (May 1996), pp. 13-- 24.

....show the necessity of mechanisms like co scheduling and migration of parallel jobs to idle workstations, but they do not predict contention effects. In fact, both [16] and [17] propose co scheduling strategies (or an approximation) for parallel applications executing in network of workstations. In [26], Harchol Balter and Downey show that preemptive migration, in which running processes may be suspended, moved to a remote host, and restarted, can minimize the effect imposed by contention on the hosts. Leutenegger and Sun [30] have created a model to show the viability of non dedicated ....

M. Harchol-Balter and A. B. Downey, "Exploiting Process Lifetime Distributions for Dynamic Load Balance", in Proceedings of ACM SIGMETRICS'96, pp. 13-24, May 1996.

....may be dropped either when the counter reaches p or number of load units requested becomes zero. Finally to answer the question, which tasks to migrate, we have decided to migrate older tasks because these tasks have a higher probability of living long enough to amortize their migration cost [7]. Age in this case refers to the CPU time a task has used thus far and not how long ago the task P2 I 3 was created as measured in wall clock time. The machine parameters used in our simulation were the same used in [8] 9] The table size was chosen to be 5 and ht, lt and ct, were 4, 10, 25 ....

....processor of the decision. To compare the relative performance of RoCLB against the other strategies, experimental simulation was used. A special purpose event driven simulator was constructed and we based our synthetic generated workloads on the work done M. Harchol Balter and A. B. Downey [7] and J. Jann et. al [14] More specifically, we model the inter arrival time using a Hyper Erlang distribution of Common Order [14] and a task s lifetime using a used better than newin expection type of distribution [7] For the other simulation parameters, such as the application s degree of ....

[Article contains additional citation context not shown here]

M. Harchol-Balter and A. B. Downey, "Exploiting process lifetime distributions for dynamic load balancing", in ACM Transactions on Computer Systems, Vol. 15, Num.

....is based on how much memory it has allocated. We can estimate application lifetimes by either maintaining a database of past application runs, or in the absence of history, we can use validated heuristics for estimating remaining lifetime based on how long a program has already been running for [HarcholBalter Downey 1997]. Finally, we record and analyze when past failures have occurred to estimate the probability of another failure occurring. Some of these variables change over the duration of the application, so we need to continually reevaluate when to checkpoint. 2 Console objects are made fault tolerant by ....

Harchol-Balter, M. and Downey, A. B. Exploiting process lifetime distributions for dynamic load balancing. ACM Transactions on Computer Systems, 15(3):253-- 285, August 1997.

....owner processes have the highest priority for their CPUs and memory allocations in their workstations, which divides the global memory system into different local regions. Regarding job migration based load sharing systems, most load sharing schemes proposed for distributed systems (e.g. 3] [6], 7] 12] mainly consider effectively sharing CPU cycles. A memory ushering algorithm is used in MOSIX for memory load sharing [2] Recently, we have developed several load sharing alternatives by considering both CPU and memory resources [11] The objective of our design is to reduce the ....

....The CPU service is switched to a different job. When page faults happen in the executions of several jobs, they will be served in FIFO order. The overhead of a remote execution is set to 0.1 second for 10Mbps Ethernet and 0. 05 for 100Mbps Ethernet based on the numbers of real systems provided by [6]. The bus service and contention are simulated based on typical bus transactions as follows. Each workstation is given a sequence number which also represents its priority rank to access the bus. The priority increases as the sequence number decreases. As multiple requests for bus services arrive ....

[Article contains additional citation context not shown here]

M. Harchol-Balter and A. B. Downey, "Exploiting process lifetime distributions for dynamic load balancing", ACM Transactions on Computer Systems, Vol. 15, No. 3, 1997, pp. 253-285.

No context found.

Mor Harchol-Balter and Allen B. Downey. Exploiting process lifetime distributions for dynamic load balancing. In Proceedings of the ACM Sigmetrics Conference on Measurement and Modeling of Computer Systems, pages 13--24, May 1996.

No context found.

Mor Harchol-Balter and Allen B. Downey. Exploiting process lifetime distributions for dynamic load balancing. Technical Report UCB//CSD-95-887, University of California at Berkeley, 1995.

No context found.

Mor Harchol-Balter and Allen Downey. Exploiting process lifetime distributions for dynamic load balancing. In Proceedings of SIGMETRICS '96, pages 13-24, 1996.

No context found.

Mor Harchol-Balter and Allen Downey. Exploiting process lifetime distributions for dynamic load balancing. In Proceedings of SIGMETRICS '96, pages 13-24, 1996.

....LF, PBS, or NQS. These schedulers typically only support run to completion (non preemptive) PS97] time divided by its service requirement. All means are per job averages. Mean slowdown is important because it is desirable that a job s delay be proportional to its processing requirement [8, 3, 13]. Users are likely to anticipate short delays for short jobs, and are likely to tolerate long delays for longer jobs. A secondary performance goal is fairness. We adopt the following definition of fairness: All jobs, long or short, should experience the same expected slowdown. In particular, long ....

....this paper we maintain the assumption that jobs are not preemptible. That is, once a job starts running, it can not be stopped and re continued where it left off. By contrast there exists considerable work on the very dif]krent problem where jobs are preemptible and maybe even migrateable, see [13] for many citations) Other generalizations of the task assignment problem include the scenario where the hosts are heterogeneous or there are multiple resources under contention. The idea of purposely unbalancing load has been suggested previously in [6] and in [4] under different contexts ....

[Article contains additional citation context not shown here]

Mor Harchol-Balter and Allen Downey. Exploiting process lifetime distributions for dynamic load balancing. ACM Trasactios o Computer S'ystems, 15(3), 1997. 26

.... Throughout this paper, whenever the distribution is mentioned, it will be assumed that is chosen such that the mean value is fixed at and the maximum value fixed at , which correspond to typical values taken from [4] Many recent measurements of computing systems [15, 9, 4, 10, 20] have observed job size distributions which are well modeled by a Bounded Pareto distribution, where . One key property of heavy tailed distributions and (many) Bounded Pareto distributions is that a tiny fraction of the very largest jobs comprise over half of the total load. We ....

Mor Harchol-Balter and Allen Downey. Exploiting process lifetime distributions for dynamic load balancing. In Proceedings of SIGMETRICS '96, pages 13--24, 1996. 21

....file requests and yet expect short response times for short requests. The slowdown of a request is therefore its response time divided by the time it would 6 require if it were the sole request in the system. Slowdown is also commonly known as normalized response time and has been widely used [10, 35, 17, 24]. ffl Mean response time as a function of request size. This will indicate whether big requests are being treated unfairly under SRPT as compared with FAIR share scheduling. 4.2 Main Experimental results (LAN) Before presenting the results of our experiments, we make some important comments. ....

M. Harchol-Balter and A. Downey. Exploiting process lifetime distributions for dynamic load balancing. ACM Transactions on Computer Systems, 15(3), 1997.

....assumptions made in our analysis via simulation, which agrees with our analysis. We evaluate our analytic formulas in the case where the job size distribution is a Bounded Pareto distribution, with a heavy tailed property. This distribution has been shown to be characteristic of computer workloads [18, 12, 7, 21]. Our PS results indicate that performance of a PS server under transient overload can be quite poor, even when mean system load is very low. The performance is dominated by the load during the overload period. The number 1 of jobs in system increases very rapidly during the overload period and ....

....we normalize the distribution by scaling down the job sizes by a factor of 3000, leaving the mean as 1. We refer to this normalized distribution as B(ff = 1:5) Pareto and Bounded Pareto distributions have been shown to be characteristic of the job size distributions in many computer workloads [18, 12, 7, 21]. These distributions have 3 important properties: 1. Decreasing failure rate (Pareto) or mostly decreasing (Bounded Pareto) 2. Infinite variance (Pareto) or very high variance (Bounded Pareto) 3. The heavy tailed property, which we define as: A very small fraction of the largest jobs (e.g. ....

Mor Harchol-Balter and Allen Downey. Exploiting process lifetime distributions for dynamic load balancing. In Proceedings of SIGMETRICS '96, pages 13--24, 1996.

....useful to know the distribution of lifetimes for all jobs. For sequential jobs on UNIX workstations, two studies have reported distributions of lifetimes with the functional form cdf L (t) P rfL tg = 1 Gamma (t=t min ) k (1) where k is a parameter that varies from workload to workload [12] [8]. When k is approximately 1, as it often is, the median remaining lifetime of a job is equal to its current age. This property is known as the past repeats heuristic. In the next section, we will show that the distribution of lifetimes for parallel scientific applications does not fit this model; ....

M. Harchol-Balter and A. B. Downey. Exploiting process lifetime distributions for dynamic load balancing. In Proceedings of the ACM Sigmetrics Conference on Measurement and Modeling of Computer Systems, pages 13--24, May 1996.

....every single request, including the very largest request, performs better under SRPT scheduling as compared with PS scheduling. The case of heavy tailed request size distributions is important because heavy tailed distributions have been shown to arise in many empirical computer workloads [19, 16, 8, 18, 24]. In particular measurements of Web file sizes and HTTP request times have been shown to be heavy tailed [6] We use the theoretical results in [21] to corroborate the results in this paper. The general idea of size based scheduling for heavy tailed workloads has also been explored in arenas ....

....large file requests and yet expect short response times for short requests. The slowdown of a request is therefore its response time divided by the time it would require if it were the sole request in the system. Slowdown is also commonly known as normalized response time and has been widely used [12, 4, 16]. Mean slowdown is the average of the slowdown of each of the requests. ffl Mean response time as a function of request size. This will indicate whether big requests are being treated unfairly under SRPT as compared with FAIR share scheduling. Before presenting the results of our experiments, we ....

M. Harchol-Balter and A. Downey. Exploiting process lifetime distributions for dynamic load balancing. ACM Transactions on Computer Systems, 15(3), 1997.

.... 2 Throughout this paper, whenever the B(k; p; ff) distribution is mentioned, it will be assumed that k is chosen such that the mean value is fixed at 3000 and the maximum value fixed at p = 10 10 , which correspond to typical values taken from [4] Many recent measurements of computing systems [15, 9, 4, 10, 20] have observed job size distributions which are well modeled by a Bounded Pareto distribution, where ff 1. One key property of heavy tailed distributions and (many) Bounded Pareto distributions is that a tiny fraction ( 1 ) of the very largest jobs comprise over half of the total load. We will ....

Mor Harchol-Balter and Allen Downey. Exploiting process lifetime distributions for dynamic load balancing. In Proceedings of SIGMETRICS '96, pages 13--24, 1996. 21

....every single request, including the very largest request, performs better under SRPT scheduling as compared with PS scheduling. The case of heavy tailed request size distributions is important because heavy tailed distributions have been shown to arise in many empirical computer workloads [21, 18, 10, 20, 24]. In particular measurements of Web file sizes and HTTP request times have been shown to be heavy tailed [8] We use the theoretical results in [7] to corroborate the results in this paper. The general idea of size based scheduling for heavy tailed workloads has also been explored in arenas other ....

....are willing to tolerate long response times for large file requests and yet expect short response times for short requests. The slowdown of a request is therefore its response time divided by its size in bytes. Slowdown is also commonly known as normalized response time and has been widely used [14, 5, 18]. Mean slowdown is the average of the slowdown of each of the requests. ffl Mean response time as a function of request size. This will indicate whether big requests are being treated unfairly under SRPT as compared with FAIR share scheduling. ffl Coefficient of variation of response time as a ....

M. Harchol-Balter and A. Downey. Exploiting process lifetime distributions for dynamic load balancing. ACM Transactions on Computer Systems, 15(3), 1997.

....fraction ( 1 ) of the very largest jobs comprise over half of the total load. We will refer to this as the heavy tailed property throughout this paper. Heavy tailed distributions with ff 1 exhibit the heavy tailed property above and appear to fit many recent measurements of computing systems [13, 7, 3, 8, 19]. Observe that lower ff values indicate greater variability in the job size distribution and stronger heavy tailed properties. In fact ff 2 indicates infinite variance. Thus, the above measurements indicate very high variability in job service requirements and a very heavy tail. In practice, ....

Mor Harchol-Balter and Allen Downey. Exploiting process lifetime distributions for dynamic load balancing. In Proceedings of SIGMETRICS '96, pages 13--24, 1996.

....distribution function F (x) PrfX xg = 1 (k=x) 3 Distribution of process lifetimes (log plot) fraction of processes with duration T) Duration (T secs. 1 1 2 1 4 1 8 1 16 1 32 1 64 1 2 4 8 16 32 64 Figure 2: Measured distribution of UNIX process CPU lifetimes, taken from [5]. Data indicates fraction of jobs whose CPU sevice demands exceed T seconds, as a function of T . A set of task sizes following a heavy tailed distribution has the following properties: 1. Decreasing failure rate: In particular, the longer a task has run, the longer it is expected to continue ....

....and the more pronounced is the heavy tailed property, i.e. the smaller the faction of large tasks that comprise half the load. As a concrete example, Figure 2 depicts graphically on a log log plot the measured distribution of CPU requirements of over a million UNIX processes, taken from paper [5]. This distribution closely ts the curve PrfProcess Lifetime Tg = 1=T: In [5] it is shown that this distribution is present in a variety of computing environments, including instructional, reasearch, and administrative environments. In fact, heavy tailed distributions appear to t many ....

[Article contains additional citation context not shown here]

Mor Harchol-Balter and Allen Downey. Exploiting process lifetime distributions for dynamic load balancing. ACM Transactions on Computer Systems, 15(3), 1997.

....how to do analysis under heavy tailed job size distributions. We then show that our analysis leads us to policies whose performance improves greatly over commonly used solutions, in some cases by orders of magnitude. This paper is a compilation of a sequence of papers which the author co wrote: [9, 6, 7, 8, 1]. Far more detail is contained in the original papers. Supported by the NSF Postdoctoral Fellowship in the Mathematical Sciences. 1 Introduction Heavy tailed distributions (also known as power law distributions) have been observed in many natural phenomena including both physical and ....

....For each problem above, we solve the problem first in the context of an exponentially distributed job size distribution and then in the context of a heavy tailed job size distribution, obtaining very different answers. This paper is based on a sequence of papers which the author co wrote: [9, 6, 7, 8, 1]. This paper only provides brief summaries of the insights learned in the original papers. Throughout the reader is referred to the original papers for more detail. 2 Measurements of Heavy tailed job size distributions in computer systems Figure 1 depicts graphically on a log log plot the ....

[Article contains additional citation context not shown here]

M. Harchol-Balter and A. Downey. Exploiting process lifetime distributions for dynamic load balancing. ACM Transactions on Computer Systems, 15(3), 1997. 15

....LSF, PBS, or NQS. These schedulers typically only support run to completion (non preemptive) PS97] time divided by its service requirement. All means are per job averages. Mean slowdown is important because it is desirable that a job s delay be proportional to its processing requirement [8, 3, 13]. Users are likely to anticipate short delays for short jobs, and are likely to tolerate long delays for longer jobs. A secondary performance goal is fairness. We adopt the following definition of fairness: All jobs, long or short, should experience the same expected slowdown. In particular, long ....

....this paper we maintain the assumption that jobs are not preemptible. That is, once a job starts running, it can not be stopped and re continued where it left off. By contrast there exists considerable work on the very different problem where jobs are preemptible and maybe even migrateable, see [13] for many citations) Other generalizations of the task assignment problem include the scenario where the hosts are heterogeneous or there are multiple resources under contention. The idea of purposely unbalancing load has been suggested previously in [6] and in [4] under different contexts ....

[Article contains additional citation context not shown here]

Mor Harchol-Balter and Allen Downey. Exploiting process lifetime distributions for dynamic load balancing. ACM Transactions on Computer Systems, 15(3), 1997. 26

No context found.

Mor Harchol Balter and Allen B. Downey. Exploiting Process Lifetime Distributions for Dynamic Load Balancing. In Proceedings of ACM Sigmetrics '96 Conference on Measurement and Modelling of Computer Systems, pages 1324, May 1996.

No context found.

Mor Harchol Balter and Allen B. Downey. Exploiting Process Lifetime Distributions for Dynamic Load Balancing. In Proceedings of ACM Sigmetrics '96 Conference on Measurement and Modelling of Computer Systems, May 1996.

No context found.

M. Harchol-Balter and A.B. Downey, "Exploiting Process Lifetime Distributions for Dynamic Load Balancing," Technical Report UCB/CSD-95-887, University of California at Berkely, November, 1995.

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