[Paper Review] Characterizing Co-located Datacenter Workloads: An Alibaba Case Study

Yue Cheng, Zheng Chai+1 more|arXiv (Cornell University)|Aug 8, 2018

Cloud Computing and Resource Management35 references4 citations

TL;DR

This paper analyzes Alibaba's publicly available co-located datacenter workload trace, characterizing interactions between long-running containerized production jobs and transient batch jobs across 1,313 machines. It reveals persistent issues like stragglers and resource overcommitment, while highlighting performance interference under high utilization and the need for a more coordinated global scheduler to optimize co-location in heterogeneous clusters.

ABSTRACT

Warehouse-scale cloud datacenters co-locate workloads with different and often complementary characteristics for improved resource utilization. To better understand the challenges in managing such intricate, heterogeneous workloads while providing quality-assured resource orchestration and user experience, we analyze Alibaba's co-located workload trace, the first publicly available dataset with precise information about the category of each job. Two types of workload---long-running, user-facing, containerized production jobs, and transient, highly dynamic, non-containerized, and non-production batch jobs---are running on a shared cluster of 1313 machines. Our multifaceted analysis reveals insights that we believe are useful for system designers and IT practitioners working on cluster management systems.

Motivation & Objective

• To understand the complex interactions and resource dynamics between long-running containerized jobs and transient batch jobs in a real-world warehouse-scale datacenter.
• To identify persistent challenges such as straggler behavior and inaccurate resource prediction in co-located environments.
• To evaluate the impact of resource contention and performance interference due to high utilization and workload heterogeneity.
• To assess the effectiveness of existing schedulers (Sigma and Fuxi) in managing co-located workloads and identify gaps in coordination.
• To inform system designers and IT practitioners about architectural improvements needed for efficient, co-location-optimized cluster management.

Proposed method

• Analyzes a publicly available 24-hour trace from Alibaba’s production cluster, including precise job category labels for 1,313 machines.
• Partitions the cluster into 'Batch only' and 'Co-located' regions to compare resource usage and utilization patterns.
• Measures CPU and memory utilization, CPI, and MPKI across different resource usage ranges to detect performance interference.
• Examines resource request, reservation, and actual usage patterns to identify overbooking, over-provisioning, and over-commitment behaviors.
• Evaluates workload dynamicity, straggler frequency, and interference between long-running and batch workloads.
• Proposes that a global Level-0 controller should coordinate schedulers by exposing resource heterogeneity caused by co-located workloads.

Experimental results

Research questions

• RQ1 How do resource usage patterns differ between long-running containerized jobs and transient batch jobs in co-located clusters?
• RQ2 To what extent do performance interference and straggler issues persist in modern co-located datacenter environments?
• RQ3 How does resource utilization in 'Batch only' regions compare to 'Co-located' regions, and what does this imply about scheduler awareness?
• RQ4 What is the impact of high CPU and memory utilization on performance metrics like CPI and MPKI?
• RQ5 How can a global controller improve coordination between workload-specific schedulers in a two-level cluster management system?

Key findings

• Long-running jobs are more memory-intensive, leading to higher overall cluster memory utilization than CPU utilization.
• Over-provisioning for long-running jobs creates significant opportunities for elastic over-commitment of unused resources by batch jobs.
• Straggler issues persist in batch workloads despite advances in scheduling, indicating ongoing challenges in prediction and resource management.
• Performance interference increases significantly at high resource utilization: maximum CPI and MPKI peak at 40%+ CPU and 80%+ memory usage.
• Fuxi, the batch scheduler, operates under the assumption of a homogeneous resource pool and does not account for the heterogeneity introduced by co-located long-running jobs.
• The lack of coordination between Sigma (long-running) and Fuxi (batch) schedulers suggests a need for a more integrated, co-location-aware global controller.