Directory Write Leases in the MagFS Distributed File System

Typical metadata-heavy workloads incur significant network round trip latencies for each namespace-modifying operations. Leases or delegations found in traditional network file systems like SMB or NFSv4 allow clients to cache directory-level information on a “read-only” basis and hence force a network round trip on every create, rename, or delete operation. This talk will focus on the concept of Directory Write Leases, a protocol-level enhancement made in the Maginatics File System (MagFS) to considerably speed up small file metadata-heavy workloads.

Directory Write Leases allow the client to act on behalf of the server for all file system operations in that directory. This is a powerful concept because it enables the client to locally serve namespace-modifying operations within that directory, and asynchronously propagate these operations to the server. We will talk about the semantics of this new lease state needed to preserve strong consistency guarantees in a distributed file system like MagFS. Finally, we will demonstrate that using Directory Write Leases we were able to hide a significant fraction of the network latency and bottlenecks for build workloads when compared to NFS or SMB, and were able to achieve a significant performance boost when compared to traditional leasing mechanisms.

Learning Objectives

Bottlenecks in small file metadata-heavy workloads e.g. build workloads
Consistency semantics and guarantees of Directory Write Leases
Implementation details and challenges of caching namespace modifying operations on the client