[Paper Review] Towards Transactional Load over XtreemFS

Roman Talyansky, Bernd Scheuermann+2 more|arXiv (Cornell University)|Jan 17, 2010

Advanced Data Storage Technologies9 references3 citations

TL;DR

This paper proposes a trace-based performance evaluation methodology for distributed file systems (DFS) under enterprise transactional workloads, using real IO traces from SAP applications to assess XtreemFS and Ceph. It finds XtreemFS delivers strong read performance and scalability with low TCO, but write latency remains a critical bottleneck requiring optimization via persistent storage enhancements.

ABSTRACT

We propose using trace-based assessment of the performance of distributed file systems (DFS) under transactional IO load. The assessment includes simulations and experiments using the IO traces. Our experiments suggest that DFS, and specifically XtreemFS have a good potential to support transactional IO load in distributed environments: they demonstrate good performance, high availability and scalability, while at the same time opening the way to TCO reduction.

Motivation & Objective

• To reduce the high overhead of evaluating distributed file systems (DFS) in real enterprise environments by avoiding full application deployment.
• To assess the suitability of object-based DFS like XtreemFS and Ceph for supporting transactional enterprise workloads.
• To evaluate performance characteristics—especially latency and throughput—of DFS under real enterprise IO traces.
• To identify performance bottlenecks, particularly in synchronous write operations, and guide future optimization.
• To validate the effectiveness of a trace-based simulation methodology for reproducible, scalable DFS evaluation.

Proposed method

• Captured real IO traces from SAP enterprise applications running on a production file system to model transactional workloads.
• Replayed the same IO traces on multiple target DFSs (XtreemFS, Ceph) to evaluate performance without deploying full applications.
• Simulated the distribution function of DFSs to analyze load balancing across Object Storage Devices (OSDs).
• Measured key performance metrics: read/write latency, throughput, and scalability across varying IO loads and number of OSDs.
• Used a controlled simulation environment to isolate file system behavior from application-specific tuning and configuration overhead.
• Evaluated the impact of configuration parameters such as stripe width and stripe size on performance and resource utilization.

Experimental results

Research questions

• RQ1 Can a trace-based methodology effectively evaluate DFS performance under real enterprise transactional workloads without deploying full applications?
• RQ2 How does XtreemFS perform under transactional IO workloads, particularly in terms of read latency, write latency, and throughput scalability?
• RQ3 What are the performance differences between XtreemFS and Ceph under increasing IO load and concurrent client streams?
• RQ4 To what extent can DFS parameters like stripe width and size be tuned to balance performance and resource utilization?
• RQ5 What are the key performance bottlenecks in XtreemFS for transactional workloads, especially in synchronous write operations?

Key findings

• XtreemFS achieves low read latency and high throughput, demonstrating strong performance for read-heavy transactional workloads.
• Write latency in XtreemFS remains significantly high, especially for synchronous writes, indicating a major performance bottleneck.
• XtreemFS scales well with increasing numbers of OSDs, distributing IO load effectively and maintaining low latencies under growing write workloads.
• CEPH shows excellent performance for small IO streams but degrades faster in latency and throughput as load increases compared to XtreemFS.
• The trace-based methodology enables reproducible, scalable, and efficient evaluation of DFS performance without full application deployment or tuning.
• XtreemFS and Ceph show strong potential for reducing TCO by enabling scale-out on commodity hardware with near-linear performance scaling.