src/ceph/doc/dev/bluestore.rst

   1 ===================
   2 BlueStore Internals
   3 ===================
   4
   5
   6 Small write strategies
   7 ----------------------
   8
   9 * *U*: Uncompressed write of a complete, new blob.
  10
  11   - write to new blob
  12   - kv commit
  13
  14 * *P*: Uncompressed partial write to unused region of an existing
  15   blob.
  16
  17   - write to unused chunk(s) of existing blob
  18   - kv commit
  19
  20 * *W*: WAL overwrite: commit intent to overwrite, then overwrite
  21   async.  Must be chunk_size = MAX(block_size, csum_block_size)
  22   aligned.
  23
  24   - kv commit
  25   - wal overwrite (chunk-aligned) of existing blob
  26
  27 * *N*: Uncompressed partial write to a new blob.  Initially sparsely
  28   utilized.  Future writes will either be *P* or *W*.
  29
  30   - write into a new (sparse) blob
  31   - kv commit
  32
  33 * *R+W*: Read partial chunk, then to WAL overwrite.
  34
  35   - read (out to chunk boundaries)
  36   - kv commit
  37   - wal overwrite (chunk-aligned) of existing blob
  38
  39 * *C*: Compress data, write to new blob.
  40
  41   - compress and write to new blob
  42   - kv commit
  43
  44 Possible future modes
  45 ---------------------
  46
  47 * *F*: Fragment lextent space by writing small piece of data into a
  48   piecemeal blob (that collects random, noncontiguous bits of data we
  49   need to write).
  50
  51   - write to a piecemeal blob (min_alloc_size or larger, but we use just one block of it)
  52   - kv commit
  53
  54 * *X*: WAL read/modify/write on a single block (like legacy
  55   bluestore).  No checksum.
  56
  57   - kv commit
  58   - wal read/modify/write
  59
  60 Mapping
  61 -------
  62
  63 This very roughly maps the type of write onto what we do when we
  64 encounter a given blob.  In practice it's a bit more complicated since there
  65 might be several blobs to consider (e.g., we might be able to *W* into one or
  66 *P* into another), but it should communicate a rough idea of strategy.
  67
  68 +--------------------------+--------+--------------+-------------+--------------+---------------+
  69 |                          | raw    | raw (cached) | csum (4 KB) | csum (16 KB) | comp (128 KB) |
  70 +--------------------------+--------+--------------+-------------+--------------+---------------+
  71 | 128+ KB (over)write      | U      | U            | U           | U            | C             |
  72 +--------------------------+--------+--------------+-------------+--------------+---------------+
  73 | 64 KB (over)write        | U      | U            | U           | U            | U or C        |
  74 +--------------------------+--------+--------------+-------------+--------------+---------------+
  75 | 4 KB overwrite           | W      | P | W        | P | W       | P | R+W      | P | N (F?)    |
  76 +--------------------------+--------+--------------+-------------+--------------+---------------+
  77 | 100 byte overwrite       | R+W    | P | W        | P | R+W     | P | R+W      | P | N (F?)    |
  78 +--------------------------+--------+--------------+-------------+--------------+---------------+
  79 | 100 byte append          | R+W    | P | W        | P | R+W     | P | R+W      | P | N (F?)    |
  80 +--------------------------+--------+--------------+-------------+--------------+---------------+
  81 +--------------------------+--------+--------------+-------------+--------------+---------------+
  82 | 4 KB clone overwrite     | P | N  | P | N        | P | N       | P | N        | N (F?)        |
  83 +--------------------------+--------+--------------+-------------+--------------+---------------+
  84 | 100 byte clone overwrite | P | N  | P | N        | P | N       | P | N        | N (F?)        |
  85 +--------------------------+--------+--------------+-------------+--------------+---------------+