The multiserver job queueing model
The multiserver job queueing model
复制标题
多服务器作业排队模型
DOI:
10.1007/s11134-022-09762-x
复制
发表时间:
2022
期刊:
影响因子:
1.2
通讯作者:
Harchol-Balter, Mor
中科院分区:
文献类型:
--
作者:
Harchol-Balter, Mor
A great deal of queueing theory is devoted to studying multiserver models, such as the M/G/n. A key feature of such models is that each job runs on a single server. Unfortunately this one-server-per-job model is not a good representation of today’s data centers. Almost all of today’s data center jobs occupy multiple servers simultaneously [12]. We refer to such jobs that run on multiple servers as multiserver jobs. A recent trace from Google’s Borg scheduler [12] shows that the number of servers utilized by a single job can vary by five orders of magnitude across jobs. Understanding the performance of systems with multiserver jobs is therefore of paramount importance. Figure 1 shows the multiserver job queueing model. Jobs arrive with average rate λ into a system with n homogeneous servers, where they are served in FCFS order. A job is of class i with probability pi. A job of class i requires (any) ni servers, which it occupies in parallel for Si hours, where Si is a random variable. Importantly, the size of a job of class i is ni· Si and is specified in units of server-hours. Some related models: While almost nothing is known about the performance of multiserver job queueing models, there is a cousin of this model, which we call the dropping model, which is analytically tractable under very general settings. In the dropping model, jobs which cannot immediately receive service are dropped. The dropping model exhibits a beautiful product form when job durations (the Si’s) are exponentially distributed, as shown in Arthurs and Kaufman [1]. Whitt [15] generalized the model to allow jobs to demand multiple resource types, while van Dijk [13] allowed durations to be generally-distributed. Related to dropping models are streaming models, which come up in communication networks. Here the resource being shared is bandwidth in the network. The “jobs” are audio or video flows which require a fixed bandwidth reservation to run (akin to needing a fixed number of servers). The goal is to schedule flows to minimize a cost related to dropping probabilities [4, 10].
登录
查看更多内容
DOI:
--
发表时间:
1987
期刊:
影响因子:
--
作者:
F. Baccelli;C. Courcoubetis;M. Reiman
通讯作者:
M. Reiman
影响因子:
5.4
作者:
P. H. Brill;L. Green
通讯作者:
L. Green
影响因子:
4.8
作者:
A. Rumyantsev;E. Morozov
通讯作者:
E. Morozov
影响因子:
1.1
作者:
N. Dijk
通讯作者:
N. Dijk
DOI:
10.1145/3492866.3549717
发表时间:
2021-09
期刊:
Proceedings of the Twenty-Third International Symposium on Theory, Algorithmic Foundations, and Protocol Design for Mobile Networks and Mobile Computing
影响因子:
--
作者:
Yige Hong;Weina Wang
通讯作者:
Yige Hong;Weina Wang