Polling Based Per-core Workqueue Management in XFS Journaling

In this paper, we analyze the scalability of XFS journaling and improve the many-core scalability of XFS journaling. We found that the lock contention in the async and await mechanisms, such as workqueue and waitqueue, is one of the main causes for the scalability failure in XFS journaling. We propose per-core pool workqueue and polling based on-disk logging to solve these problems. By using per-core pool workqueue, we resolve the lock contention on thread pool used in the workqueue module. By using polling based on-disk logging, XFS waits for the journal thread to finish in polling based mechanism. We resolve the lock contention on global waitqueue and reduce the latency of on-disk logging through polling. We implement these methods based on XFS. In varmail and exim workloads, the proposed techniques improve the benchmark performance by 57% and 28% against XFS, and 9% and 28% against ScaleXFS, respectively.