Cache in Chromium: Disk Cache

Cache in ChromiumDisk Cache & Overall cache flow

Chang W. Doh

GDG Korea WebTech OrganizerHTML5Rocks/KO Contributor/Coordinator

</hi><hi>

Before stepping into Cache:Network stack

● A mostly single-threaded cross-platform library primarily for resource fetching

○ URLRequest■ represents the request for a URL

○ URLRequestContext■ contains all the associated contexts to fullfill

the ‘URL request’● e.g. cookies, host resolver, cache

Network Stack:What’s ‘Network Stack’?

● Some code layouts○ /net/base

■ shared utilities for /net modules○ /net/disk_cache

■ Cache for web resources○ /net/url_request

■ URLRequest, URLRequestContext, ...

Before stepping into Cache:Network stack

Typical request-flow

HttpCache Cache(aka Disk Cache)check

HttpCache::Transaction

not exist

notify

notify

Cache hit!

Disk Cache:(a.k.a. Cache)

● Cache ○ Stores resources fetched from the web○ A part of ‘/net’

■ location: /net/disk_cache■ This means ‘DiskCache’ will controll cache-flows

for network fetches.

● NOTE:○ Android use ‘Simple cache’.

■ location: /net/disk_cache/simple

DiskCache:What’s DiskCache?

● Main characteristics:○ The cache should not grow unbounded

■ Algorithm to decide when removing old entries○ Not critical to loose some data

■ But discarding whole cache should be minimized

○ Access should be possible to use sync or async operations

○ Design should avoid ‘cache trashing’

DiskCache:Characteristics

● Main characteristics:○ Should be possible to remove a entry from

the cache■ and keep working with that entry while at the

same time inaccessible to other requests○ Shouldn’t be using explicit multithread sync

■ Always called from the same thread

■ However, callbacks must be issued by message loop for avoiding reentrancy

DiskCache:Characteristics

● /net/disk_cache/disk_cache.h

● 2 Interfaces○ disk_cache::Backend

■ manages entries on the cache○ disk_cache::Entry

■ handles operations specific to a given resource

DiskCache:External interfaces

● An entry is identified by its key

○ e.g. http://www.google.com/favicon.ico

● Once an entry is created, the data is stored in separate chunks or data streams:

○ HTTP headers

○ Actual resource data

● Index for the required stream is an argument to methods:

○ Entry::ReadData

○ Entry::WriteData

External interfaces:Backend

Very Simple Cache(a.k.a. Simple Cache)

● Proposed to a new backend for diskcache○ Conforming to the interface in Disk Cache○ Very simple

■ Using 1 file per cache entry + index file■ Dealing with I/O bottlenecks

Simple cache:What is “Simple Cache”?

Comparison toBlockfile Backend

?

● Comparison to blockfile cache○ More resilent under corruption from the

system crash■ Periodcally flushes its entire index■ Swaps index in atomically

■ After system crash, will starts with the stale cache

● NOTE: With the blockfile cache, chrome will drops whole cache by default

Simple cache:Benefits and goals

● Comparison to blockfile cache (cont’d)○ Doesn’t delay launching network requests

■ Elimination of delay factors● No context switching● Not blocks disk I/O before using network

■ Blockfile has (AVR) 14~25ms delay on requests

● On Android, slower flash controllers make these delays significantly slower.

Simple cache:Benefits and goals

● Comparison to blockfile cache (cont’d)○ Lower resident set pressure & fewer IO ops.

■ Disk format has● 256~512B per entry records● + rankings & index information(~100B) per entry● Not all entries that are heavily used contiguously

■ Simple cache● stores only SMALLNUM bytes per entry in memory● doesn’t access the disk where not required

Simple cache:Benefits and goals

● Comparison to blockfile cache (cont’d)○ Simpler

■ Shorter and easier via explicitly avoiding implementation of filesystem than blockfile’s

Simple cache:Benefits and goals

● Not a log structed cache system○ I/O performs by Simple cache is mostly

sequential. But NOT log structed■ If it is, it means “filesystem that itself is log

structed.”

● Not a filesystem○ Disk cache delegates filesystem.

■ means “Simple cache uses abstract interface of Disk Cache instead implementing its own filesystem”.

Simple cache:Non-goals; Simple cache is

● Entry hash○ Hash with 40 bit SHA-2 of url○ 2 entries with same EH can’t be stored

● Stored in single directory○ ONE index file○ Each entry stored in a single file

■ named by HexEntryHash_StreamNumber

Simple cache:Structure on Disk

● A file ‘00index’○ contains data for initializing memory index

● Index (on memory)○ used for faster cache performance○ consists of entry hashes for records & simple

eviction information

Simple cache:Structure on Disk

● Formats of entry file○ Simple file header

○ Simple file EOF

○ Simple File sparse range header

Simple cache:Structure on Disk

magic_number version key_length key_hash

flagsfinal_magic_number data_crc32 stream_size

sparse_magic_number offset length data_crc32

Term:Sparse file

● I/O thread operations○ public API is called on the I/O thread○ The index is updated in the I/O thread

● Worker pool operations○ All I/O operations are performed async on

the worker pool.○ Cache will keep a pool of new entries ready

to move into final place.

Simple cache:Implementation

● Index flushing & consistency checking○ The index is flushed on

■ shutdown■ periodically

● Operation without index○ can operate without the IO thread index by

directly opening files in the directory.○ for avoiding startup speeds & I/O costs

Simple cache:Implementation

[1] Disk Cache[2] Disk Cache 3.0[3] Very Simple Cache[4] Multi-process Resource Loading[5] Network Stack[6] Network Stack Use in Chromium

References