The Power of d Choices in Scheduling for Data Centers with Heterogeneous Servers

TL;DR

This paper explores simplified, low-complexity load balancing algorithms using the power of $d$ choices in heterogeneous data center servers, demonstrating improved performance and efficiency over traditional methods.

Contribution

It introduces the Balanced-Pandas-Pod algorithm, combining $d$ choices with Balanced-Pandas, showing better performance and lower complexity than existing algorithms in large-scale data centers.

Findings

Balanced-Pandas-Pod outperforms simple Balanced-Pandas at low and medium loads.

Balanced-Pandas-Pod performs nearly as well as Balanced-Pandas at high loads.

Complexity of proposed algorithms is reduced to O(1), enabling faster scheduling and energy savings.

Abstract

MapReduce framework is the de facto in big data and its applications where a big data-set is split into small data chunks that are replicated on different servers among thousands of servers. The heterogeneous server structure of the system makes the scheduling much harder than scheduling for systems with homogeneous servers. Throughput optimality of the system on one hand and delay optimality on the other hand creates a dilemma for assigning tasks to servers. The JSQ-MaxWeight and Balanced-Pandas algorithms are the states of the arts algorithms with theoretical guarantees on throughput and delay optimality for systems with two and three levels of data locality. However, the scheduling complexity of these two algorithms are way too much. Hence, we use the power of $d$ choices algorithm combined with the Balanced-Pandas algorithm and the JSQ-MaxWeight algorithm, and compare the complexity of the simple algorithms and the power of $d$ choices versions of them. We will further show that the Balanced-Pandas algorithm combined with the power of the $d$ choices, Balanced-Pandas-Pod, not only performs better than simple Balanced-Pandas, but also is less sensitive to the parameter $d$ than the combination of the JSQ-MaxWeight algorithm and the power of the $d$ choices, JSQ-MaxWeight-Pod. In fact in our extensive simulation results, the Balanced-Pandas-Pod algorithm is performing better than the simple Balanced-Pandas algorithm in low and medium loads, where data centers are usually performing at, and performs almost the same as the Balanced-Pandas algorithm at high loads. Note that the load balancing complexity of Balanced-Pandas and JSQ-MaxWeight algorithms are $O(M)$, where $M$ is the number of servers in the system which is in the order of thousands servers, whereas the complexity of Balanced-Pandas-Pod and JSQ-MaxWeight-Pod are $O(1)$, that makes the central scheduler faster and saves energy.