Very Simple Backend
Proposed is a new backend for the disk cache, conforming to the interface in Disk Cache. The new backend is purposefully very simple, using one file per cache entry, plus an index file. This backend will be useful as a testing baseline, as well as dealing with IO bottlenecks impairing mobile browsing performance on some platforms.
Compared to the standard blockfile cache, this new design has benefits and goals:
It is more resilient under corruption from system crashes. The new design periodically flushes its entire index, and swaps it in atomically. This single atomic operation, together with cache entries being in separate files similarly swapped in makes the implications of system crash much less serious; after system crash Chrome will start with a stale cache, rather than have to drop its entire cache as many users experience with the blockfile cache. It does not delay launching network requests. The current cache requires multiple context switches, and possibly blocking disk IO before launching requests that ultimately use the network. These delays average over 25ms on requests using the network, and about 14ms averaged over all requests on Windows. Without context switches or blocking disk IO, this can be entirely eliminated. On the Android platform, the slower flash controllers make these delays significantly slower, increasing this benefit of a very simple backend. Lower resident set pressure and fewer IO operations. Our disk format has 256-512byte per entry records, plus rankings & index information of ~100bytes per entry in resident set pressure. Not all entries that are heavily used are contiguous, so paging pressure is in practice larger. The very simple cache stores only SMALLNUM bytes per entry in memory, contiguously, and does not normally access the disk in the critical path of a request where not required. As a result, the Very Simple Backend should maintain good performance even without good OS buffer cache.
Simpler. The very simple cache explicitly avoids implementing a filesystem in chrome. Files are opened and either read or written from the beginning to end. No reallocations within files take place. The cache thread component of the very simple cache should be shorter and easier to maintain than the filesystem-in-chrome approach. Together with the above benefits and goals, the new design has some non-goals: It is not a log structured cache. While the IO performed by the Very Simple Cache is mostly sequential, it is not fundamentally log structured; in particular, the filesystem operations it performs are not log structured unless used on a filesystem that itself is log structured. It is not a filesystem. The disk cache delegates filesystem operations to the filesystem. As filesystems improve, or change for different devices, the disk cache will benefit simultaneously.
See Bug 173381 for status on this implementation, or track the Internals-Network-Cache-Simple issue in crbug. For related designs covering performance tracking, see Disk Cache Benchmarking & Performance Tracking.
Like the blockfile disk cache, the disk cache is stored in a single directory. There is one index file, and each entry is stored in a single file in that directory. An Entry Hash is a relatively short hash of the url, used for storage efficiency in index and entry naming. Two entries with the same Entry Hash cannot be stored. With a 40 bit per-user-salted SHA-2 of the url; collisions would occur only at one in a million probability with a million entry cache. Each cache entry is stored in a file named by the Entry Hash in hexadecimal, an underscore, and the backend stream number. The cache is stored in a single directory, with a file file 00index that contains data for initializing an in memory index used for faster cache performance. The index consists of entry hashes for records, together with simple eviction information. Format of Entry Files:
- A magic number and version field.
- The full url, length prefixed.
- Chunks of data, length prefixed, compressed and checksummed.
- An EOF record.
The public API is called on the IO thread; and the cache maintains the index data on the IO thread as well. This allows for cache misses, and cache create operations to be performed at low latency. The index is updated in the IO thread as well.
All IO operations in the simple cache are performed asynchronously on a worker thread pool.
To reduce the critical path cost of creating new entries, the simple cache will keep a pool of newly created entries ready to move into final place; this permits the IO thread to update its index and provide a new entry at zero latency, with some cost in renames-into-place occurring later in the entries life cycle.
The index is flushed on shutdown, and periodically while running to guard against system crashes. On recovery from a system crash, the browser will have a stale index, and thus will need to periodically iterate over the cache directory to find entries in the directory not mentioned in the idex.
If startup speeds and startup IO is too costly, note that the simple backend can operate without the IO thread index by directly opening files in the directory.
Request conditionalization, freshness lifetime, and HTTP validator information could be stored in the index on the IO thread, greatly reducing latency required to serve requests.
It's arguably not a very simple idea, but in practice combining the headers and the entry data into the same file makes atomic filesystem operations (renames, etc...) easier, while also making writes/reads of a single entry generally more sequential.