Phoenix: A Constraint-aware Scheduler for Heterogeneous Datacenters

[gaocegege]

http://ieeexplore.ieee.org/abstract/document/7980039/

http://www.cse.psu.edu/hpcl/docs/2017_ICDCS_Prashanth.pdf

[at15]

.w.

[gaocegege]

这篇文章是讲如何设计一个对 Constraint 友好的调度器的. 在这里 Constraint 概念就是用户额外书写的对 QoS, 或者任务运行环境的一些个性化设定, 比如要跑在 GPU 或者 FPGA 上等等.

它最大的贡献是提出了一个 Constraint Resource Vector 的概念:

The CRV of a node is a vector of node resources represented as <cpu, mem, disk, os, clock, net_bandwidth>. For every node, the CRV_Monitor calculates the ratio of demand and supply for every constraint per heartbeat interval and updates the CRV_Lookup_Table.

Using this information we further estimate the waiting time (E[W]) for every queue using Pollaczek-Khinchin M/G/1 queuing theory model based estimator.

它们用了排队论来估计等待时间, 并且基于 Constraint Resource Vector 去做任务的重排序. 最后它的 evaluation 是用 simulation 的方式做的, 虽然用了三个 trace, 但是我还是对它持怀疑态度.