9srv Manual Collection/plan9/venti(8) | 9srv Manual Collection/plan9/venti(8) |
---|
A venti server requires multiple disks or disk partitions, each of which must be properly formatted before the server can be run.
The data log is logically split into sections called arenas, typically sized for easy offline backup (e.g., 500MB). A data log may comprise many disks, each storing one or more arenas. Such disks are called arena partitions. Arena partitions are filled in the order given in the configuration.
The index is logically split into block-sized pieces called buckets, each of which is responsible for a particular range of scores. An index may be split across many disks, each storing many buckets. Such disks are called index sections.
The index must be sized so that no bucket is full. When a bucket fills, the server must be shut down and the index made larger. Since scores appear random, each bucket will contain approximately the same number of entries. Index entries are 40 bytes long. Assuming that a typical block being written to the server is 8192 bytes and compresses to 4096 bytes, the active index is expected to be about 1% of the active data log. Storing smaller blocks increases the relative index footprint; storing larger blocks decreases it. To allow variation in both block size and the random distribution of scores to buckets, the suggested index size is 5% of the active data log.
The (optional) bloom filter is a large bitmap that is stored on disk but
also kept completely in memory while the venti server runs.
It helps the venti server efficiently detect scores that are
not
already stored in the index.
The bloom filter starts out zeroed.
Each score recorded in the bloom filter is hashed to choose
nhash
bits to set in the bloom filter.
A score is definitely not stored in the index of any of its
nhash
bits are not set.
The bloom filter thus has two parameters:
nhash
(maximum 32)
and the total bitmap size
(maximum 512MB, 2u32d bits).
The bloom filter should be sized so that
nhash
×
nblock
<=
0.7 ×
b,
where
nblock
is the expected number of blocks stored on the server
and
b
is the bitmap size in bits.
The false positive rate of the bloom filter when sized
this way is approximately 2u-nblockd.
Nhash
less than 10 are not very useful;
nhash
greater than 24 are probably a waste of memory.
Fmtbloom
(see
venti-fmt(8))
can be given either
nhash
or
nblock;
if given
nblock,
it will derive an appropriate
nhash.
The lump cache holds recently-accessed venti data blocks, which the server refers to as lumps. The lump cache should be at least 1MB but can profitably be much larger. The lump cache can be thought of as the level-1 cache: read requests handled by the lump cache can be served instantly.
The block cache holds recently-accessed disk blocks from the arena partitions. The block cache needs to be able to simultaneously hold two blocks from each arena plus four blocks for the currently-filling arena. The block cache can be thought of as the level-2 cache: read requests handled by the block cache are slower than those handled by the lump cache, since the lump data must be extracted from the raw disk blocks and possibly decompressed, but no disk accesses are necessary.
The index cache holds recently-accessed or prefetched index entries. The index cache needs to be able to hold index entries for three or four arenas, at least, in order for prefetching to work properly. Each index entry is 50 bytes. Assuming 500MB arenas of 128,000 blocks that are 4096 bytes each after compression, the minimum index cache size is about 6MB. The index cache can be thought of as the level-3 cache: read requests handled by the index cache must still go to disk to fetch the arena blocks, but the costly random access to the index is avoided.
The size of the index cache determines how long venti can sustain its `burst' write throughput, during which time the only disk accesses on the critical path are sequential writes to the arena partitions. For example, if you want to be able to sustain 10MB/s for an hour, you need enough index cache to hold entries for 36GB of blocks. Assuming 8192-byte blocks, you need room for almost five million index entries. Since index entries are 50 bytes each, you need 250MB of index cache. If the background index update process can make a single pass through the index in an hour, which is possible, then you can sustain the 10MB/s indefinitely (at least until the arenas are all filled).
The bloom filter requires memory equal to its size on disk, as discussed above.
A reasonable starting allocation is to divide memory equally (in thirds) between the bloom filter, the index cache, and the lump and block caches; the third of memory allocated to the lump and block caches should be split unevenly, with more (say, two thirds) going to the block cache.
The venti web server provides the following URLs for accessing status information:
Requests for other files are served by consulting a directory named in the configuration file (see webroot below).
The configuration file consists of lines in the form described below. Lines starting with # are comments.
After formatting a venti system using fmtindex, the order of arenas and index sections should not be changed. Additional arenas can be appended to the configuration; run fmtindex with the -a flag to update the index.
The configuration file also holds configuration parameters for the venti server itself. These are:
The units for the various cache sizes above can be specified by appending a k, m, or g (case-insensitive) to indicate kilobytes, megabytes, or gigabytes respectively.
The file name in the configuration lines above can be of the form file:lo-hi to specify a range of the file. Lo and hi are specified in bytes but can have the usual k, m, or g suffixes. Either lo or hi may be omitted. This notation eliminates the need to partition raw disks on non-Plan 9 systems.
Format the index sections, the arena partition, the bloom filter, and finally the main index:
Start the server and check the storage statistics:
9srv Manual Collection/plan9/venti(8) | Rev: Mon Feb 18 20:00:48 GMT 2013 |