Who limits the resource efficiency of my datacenter: An analysis of Alibaba datacenter traces

Trace analysis in Alibaba production clusters, which co-locates different workloads to improve resource efficiency.

Metadata

Presented in IWQoS 2019.

Authors: Jing Guo, Zihao Chang, Sa Wang, Haiyang Ding, Yihui Feng, Liang Mao, Yungang Bao (ICT, CAS & Alibaba)

Trace: https://github.com/alibaba/clusterdata/tree/v2018

Understanding the paper

This paper analyzes the Alibaba production trace, which co-locates different workloads to improve resource efficiency.

Trace: Homogeneous cluster. Each server with 96 cores and 1 unit memory normalized.

Key insights

1. Memory becomes the new bottleneck of datacenters => require efficient memory reclaim techniques

2. Batch-processing applications are treated as second-class citizens and restricted to utilize limited resources => potentially overprotect latency-critical applications

3. More than 90% of latency-critical applications are written in Java => massive self-contained JVMs complicate resource management

Questions

There are many Java applications in Alibaba, which may cause these issues. Is this a common scenario?