LMDB keyvalue backend for Ceph » History » Version 1
Jessica Mack, 07/10/2015 05:36 PM
| 1 | 1 | Jessica Mack | h1. LMDB keyvalue backend for Ceph |
|---|---|---|---|
| 2 | |||
| 3 | h3. Summary |
||
| 4 | |||
| 5 | We have completed performance evaluation of current key/value store with rocksdb and leveldb, but the performance is worse than filestore, especially for 4K random cases, the major reason is due to impaction caused by the LSM-way implementation. LMDB is an ultra-fast, ultra-compact key-value embedded data store developed by Symas for the OpenLDAP Project. It uses memory-mapped files, so it has the read performance of a pure in-memory database while still offering the persistence of standard disk-based databases, and is only limited to the size of the virtual address space |
||
| 6 | |||
| 7 | refer to #11028 if you want to check progress |
||
| 8 | |||
| 9 | h3. Owners |
||
| 10 | |||
| 11 | * Xinxin Shu (Intel) |
||
| 12 | * Jian Zhang (Intel) |
||
| 13 | * Name |
||
| 14 | |||
| 15 | h3. Interested Parties |
||
| 16 | |||
| 17 | * Name (Affiliation) |
||
| 18 | * Name (Affiliation) |
||
| 19 | * Name |
||
| 20 | |||
| 21 | h3. Current Status |
||
| 22 | |||
| 23 | KeyValueStore performance is Lower than Filestore performance in Seq.Wr, Rand.Wr and Rand.Rd |
||
| 24 | * Rocksdb compared with Filestore achieves 17% (Seq.Wr), 63%(Rand.Wr) and 54% (Rand.Rd) |
||
| 25 | * Leveldb compared with Filestore achieves 23% (Seq.Wr), 22%(Rand.Wr) and 26% (Rand.Rd) |
||
| 26 | |||
| 27 | KeyValueStore seq.wr performance is throttled by its backend write compaction mechanism, where the SSD utilization is near 90%. |
||
| 28 | |||
| 29 | h3. Detailed Description |
||
| 30 | |||
| 31 | Current filestore is not capable of fully utilize full SSD setup's capacity, and considering currently there are three backend store: memstore, filestore, key-value store, we are exploring whether the key/value store could help in the full SSD setup's scenario. |
||
| 32 | LMDB is a read-optimized design and performs reads several times faster than other DB engines, several orders of magnitude faster in many cases. It is not a write-optimized design, please refer to ondisk microbench http://symas.com/mdb/ondisk/. since random get is relative slow for rocksdb & leveldb, it would be worthwhile to add LMDB as a new key/value backend store for Ceph. |
||
| 33 | |||
| 34 | h3. Work items |
||
| 35 | |||
| 36 | h4. Coding tasks |
||
| 37 | |||
| 38 | # Task 1 |
||
| 39 | # Task 2 |
||
| 40 | # Task 3 |
||
| 41 | |||
| 42 | h4. Build / release tasks |
||
| 43 | |||
| 44 | # Task 1 |
||
| 45 | # Task 2 |
||
| 46 | # Task 3 |
||
| 47 | |||
| 48 | h4. Documentation tasks |
||
| 49 | |||
| 50 | # Task 1 |
||
| 51 | # Task 2 |
||
| 52 | # Task 3 |
||
| 53 | |||
| 54 | h4. Deprecation tasks |
||
| 55 | |||
| 56 | # Task 1 |
||
| 57 | # Task 2 |
||
| 58 | # Task 3 |