The Metadata Problem
Unruly NFS Metadata Problem: Metadata Performance Tax
Processing NFS metadata requests can constitute a serious performance bottleneck. Clients and servers making use of NFS, exchange a great deal of information about files, information that is referred to as NFS metadata, before files can be read from or written to. It doesn’t matter whether a client requests a file or browses a directory, numerous metadata operations must occur first. That metadata conversation is cycle intensive and can seriously impede NAS system performance. Here’s why:
- NFS is a very ‘chatty” stateless protocol that requires a good deal of informational exchange between clients and servers to communicate the status of a file, including denoting when a file was last modified, permission accesses and pathnames stored in symbolic links.
- Each metadata conversation adds multiple round trips between each NFS client and the NAS system. This can translate into hundreds of thousands to millions of executed NFS metadata operations per second.
- NFS metadata operations can easily make up 80 to 90% of the total NFS operations on a given NAS system especially when work on a file occurs concurrently in parallel among several clients. This is something that is common in the film and video post production work, life science genome and protein decoding, pharmaceutical molecular analysis, seismic downstream processing and software enginerering.
- Workflow sharing requires each client to know the status of files it is accessing sharing. Ensuring coherency requires reading statuses across hundreds if not thousands of clients for an enormous number of files, creating a great deal of traffic for the NAS server and severely impacting performance. The user experiences this as increased response times.
Processing large numbers of NFS metadata requests consumes too much NAS controller resources. Which in turn, reduces the NAS system’s capability to provide its most critical function of serving payload data to clients. Metadata processing can lead to processor induced excessive latency especially as the amount of NFS metadata escalates. Response times slow to a crawl.