]> git-server-git.apps.pok.os.sepia.ceph.com Git - ceph.git/commitdiff
doc/dev/bluestore: write path notes
authorSage Weil <sage@redhat.com>
Sat, 14 May 2016 12:23:15 +0000 (08:23 -0400)
committerSage Weil <sage@redhat.com>
Wed, 1 Jun 2016 15:38:48 +0000 (11:38 -0400)
Signed-off-by: Sage Weil <sage@redhat.com>
doc/dev/bluestore.rst [new file with mode: 0644]

diff --git a/doc/dev/bluestore.rst b/doc/dev/bluestore.rst
new file mode 100644 (file)
index 0000000..7568773
--- /dev/null
@@ -0,0 +1,85 @@
+===================
+BlueStore Internals
+===================
+
+
+Small write strategies
+----------------------
+
+* *U*: Uncompressed write of a complete, new blob.
+
+  - write to new blob
+  - kv commit
+
+* *P*: Uncompressed partial write to unused region of an existing
+  blob.
+
+  - write to unused chunk(s) of existing blob
+  - kv commit
+
+* *W*: WAL overwrite: commit intent to overwrite, then overwrite
+  async.  Must be chunk_size = MAX(block_size, csum_block_size)
+  aligned.
+
+  - kv commit
+  - wal overwrite (chunk-aligned) of existing blob
+
+* *N*: Uncompressed partial write to a new blob.  Initially sparsely
+  utilized.  Future writes will either be *P* or *W*.
+
+  - write into a new (sparse) blob
+  - kv commit
+
+* *R+W*: Read partial chunk, then to WAL overwrite.
+
+  - read (out to chunk boundaries)
+  - kv commit
+  - wal overwrite (chunk-aligned) of existing blob
+
+* *C*: Compress data, write to new blob.
+
+  - compress and write to new blob
+  - kv commit
+
+Possible future modes
+---------------------
+
+* *F*: Fragment lextent space by writing small piece of data into a
+  piecemeal blob (that collects random, noncontiguous bits of data we
+  need to write).
+
+  - write to a piecemeal blob (min_alloc_size or larger, but we use just one block of it)
+  - kv commit
+
+* *X*: WAL read/modify/write on a single block (like legacy
+  bluestore).  No checksum.
+
+  - kv commit
+  - wal read/modify/write
+
+Mapping
+-------
+
+This very roughly maps the type of write onto what we do when we
+encounter a given blob.  In practice it's a bit more complicated since there
+might be several blobs to consider (e.g., we might be able to *W* into one or
+*P* into another), but it should communicate a rough idea of strategy.
+
++--------------------------+--------+--------------+-------------+--------------+---------------+
+|                          | raw    | raw (cached) | csum (4 KB) | csum (16 KB) | comp (128 KB) |
++--------------------------+--------+--------------+-------------+--------------+---------------+
+| 128+ KB (over)write      | U      | U            | U           | U            | C             |
++--------------------------+--------+--------------+-------------+--------------+---------------+
+| 64 KB (over)write        | U      | U            | U           | U            | U or C        |
++--------------------------+--------+--------------+-------------+--------------+---------------+
+| 4 KB overwrite           | W      | W            | W           | R+W          | P|N (F?)      |
++--------------------------+--------+--------------+-------------+--------------+---------------+
+| 100 byte overwrite       | R+W    | W            | R+W         | R+W          | P|N (F?)      |
++--------------------------+--------+--------------+-------------+--------------+---------------+
+| 100 byte append          | R+W    | W            | R+W         | R+W          | P|N (F?)      |
++--------------------------+--------+--------------+-------------+--------------+---------------+
++--------------------------+--------+--------------+-------------+--------------+---------------+
+| 4 KB clone overwrite     | P|N    | P|N          | P|N         | P|N          | N (F?)        |
++--------------------------+--------+--------------+-------------+--------------+---------------+
+| 100 byte clone overwrite | P|N    | P|N          | P|N         | P|N          | N (F?)        |
++--------------------------+--------+--------------+-------------+--------------+---------------+