PAGE 5
Figure 2: SanDisk Cache/Store Capacity Increase through Flash Subsystem
Figure 3 shows the throughput of Membrain relative to Membase in single node and replicated
2 node configurations. The test was performed with dual-socket Westmere servers with 1/2 TB
of flash memory. The workload has an average object size of 2kB executing 90% single gets
and 10% puts. Membrain achieves over 5 times the throughput in single node and replicated
2 node configurations.
Figure 3: Throughput Comparisons of Membrain vs Membase
Membrain delivers large consolidation relative to alternative NoSQL key-value stores, since
their performance drops dramatically when data does not fit in the server DRAM. This is shown
in Figure 4 which compares the single-node performance of Membrain, SanDiskSQL
™
(a full
high-availability build of MySQL and its InnoDB storage engine), Cassandra, and MongoDB.
Each is executing the same key-value random query benchmark of 32 million and 64 million
1 KB items, on the same hardware (dual quad-core Intel processors with 64 GB of DRAM and
512GB of flash memory). The in-memory key-value random query performance of Membrain
greatly exceeds other NoSQL alternatives, giving from 2 to 100 times improvement. The per-
formance of Membrain when data cannot fit in DRAM is at least 25 times that of the alternative
key-value stores.
Figure 4: Throughput Comparisons of Key Value Store Alternatives
Cache capacity per node (GB)
100
200
0
300
400
500
600
SanDisk
Traditional Memcached
50,000
100,000
0
150,000
200,000
250,000
300,000
Membase
1.6.0.1
SanDisk
Membrain 3.0
Membase
1.6.0.1
SanDisk
Membrain 3.0
Single Node
Persistent Key Value Store
Replicated
Key Value Store
in DRAM
in Flash
Cassandra
10,500
1,790
MongoDB
49,000
4,000
Membrain
310,000
160,000