§50.6–50.9AFS Crash Recovery, Scale and Performance, and Summary

We’ve seen how AFS scales; now the costs and the payoff.

50.6 Crash Recovery

Because callbacks are state, AFS crash recovery is more involved than NFS’s:

The upshot: a server crash is a big event — every client must be told to distrust its cache — whereas an NFS client hardly noticed. That’s the price AFS pays for a more scalable and understandable caching model.

50.7 Scale and Performance

With callbacks in place, AFSv2 was measured to scale to about 50 clients per server (up from 20), and client performance often came close to purely local, since in the common case every access is served from the local cache.

To compare with NFS, we reason analytically. Assume files of $N_s$, $N_m$, or $N_L$ blocks (small/medium/large, with only large files exceeding client memory); a remote block access costs $L_{net}$, local disk $L_{disk}$, local memory $L_{mem}$, and $L_{net} > L_{disk} > L_{mem}$:

The pattern: mostly equivalent, with two telling exceptions. AFS wins on a large-file re-read (its on-disk cache holds a file too big for NFS’s memory cache, which must re-fetch). AFS loses whenever only a small part of a large file is touched, and on overwrites — because it fetches (and re-stores) the whole file regardless.

50.8 Other Improvements

Like FFS before it, AFS used the chance to add genuinely nice features:

• A true global namespace global namespace A design, provided by AFS, in which every file is named the same way on all client machines, so a given path refers to the same file everywhere. This contrasts with NFS, where each client may mount servers at whatever mount points it chooses, making consistent naming across clients only a matter of convention and administrative effort. defined in ch. 50 — open in glossary — every file is named the same way on every client. (NFS lets each client mount servers however it likes, so consistent naming is only a convention.)
• Real security: authenticating users and keeping files private. (Early NFS security was famously weak.)
• Flexible, user-managed access-control lists, and strong management tools for administrators — areas where AFS was well ahead of its time.

50.9 Summary

AFS shows a very different path from NFS: by minimizing server interactions through whole-file caching and callbacks, one server supports many clients, and its consistency model is simple to reason about — free of NFS’s occasional weirdness. Its namespace, security, and access control were ahead of their time.

And yet AFS is on the decline: NFS became an open standard, so many vendors supported it, and — along with Windows’ CIFS — it dominates the market. AFS’s lasting influence is its ideas rather than the system itself. Tellingly, NFSv4 added server state (an open protocol message), growing ever more AFS-like. Next, a closing dialogue wraps up the distributed arc — and the book.