Improvements in Bitsy 1.5
0 likes3,243 views
Bitsy 1.5 features improvements to memory efficiency through more compact data structures and lock-free reading algorithms. Benchmarks show the read throughput exceeds 10M reads/sec and is comparable to Neo4J when the graph fits in memory. Bitsy continues to outperform in write throughput due to its "No Seek" writing principle. The release is available with AGPL or commercial licensing options.
Improvements in Bitsy 1.5
- 1. Improvements in Bitsy 1.5 Sridhar Ramachandran Founder, LambdaZen LLC
- 2. Background ● Bitsy is a small, fast, embeddable, durable, in-memory graph database that implements the Tinkerpop Blueprints API. ● The original presentation on Bitsy is available at http://slideshare.net/lambdazen/bitsy-graphdatabase ● Bitsy 1.5 is faster and leaner than before! ○ Has a smaller memory footprint ○ Uses (mostly) lock-free read algorithms ● This presentation covers the improvements in the 1.5 release.
- 3. Major features in the 1.5 release ● The 1.5 release features: ○ Memory-efficient data structures ○ Mostly lock-free read algorithms ● Bitsy’s new memory-efficient data structures are designed to reduce the overhead of maintaining adjacency lists and properties. ● Bitsy’s new read algorithms are designed to use the latest Java “compare-and-set” (CAS) concurrency features to reduce the overhead of locks in highly threaded scenarios.
- 4. Memory-efficient data structures ● Bitsy 1.0 relied on Java Collections to maintain adjacency lists and properties of vertices. ● Java Collections aren’t memory efficient for small-sized data structures because they create many holder objects. ● The 1.5 release stores small adjacency lists (N<24) and small properties (N<16) in hand- coded objects with minimal overhead.
- 5. Memory-efficient data structures ● Different concrete classes capture adjacency lists and properties for small N. ○ This approach reduces the overall number of objects. ○ Large adjacency lists are stored in a compact hash- set by label referring to memory-efficient lists. Adjacency lists for out-degree 0, 1 and 2 Vertex properties for N = 0, 1 and 2
- 6. Lock-free reading ● Bitsy 1.5 also introduces lock-free reading using sequential locks (seqlock). ● Read operations track the sequence numbers at the start and end. ○ If they are the same -- Success. ○ If they are different -- Retry! ● Reads don’t start till the counter is even. ● Writers increment the counters twice ○ Before the write to make the counter an odd number ○ After the write to make the counter an even number
- 7. (Mostly) lock-free reading ● Bitsy’s sequential locks can cause “live lock” situations when there are too many writers. ● To avoid this, readers degrade to RW locks after a certain number of retries. ● Seqlock are faster than RW locks in highly threaded environments where the # of active threads exceed the # of cores. ● Bitsy uses locks on writes because ○ write-retries are complex with transactions, and ○ locking is not the bottleneck for writes -- the file system is the bottleneck.
- 8. Benchmarks ● The plot below shows the read throughput* of a test! application that repeatedly loops through a graph. * Tests performed on a $600 HP p7-1287c desktop PC with a single 7200rpm hard disk. ! The code for this test can be found in BitsyGraphTest.java under the method testMultiThreadedCommits().
- 9. Benchmarks ● The lock-free read algorithms in Bitsy 1.5 show a significantly higher throughput than Bitsy 1.0. ○ Bitsy 1.0 had a drop in performance when the number of threads exceeded the number of cores. ○ The read throughput exceeds 10M reads/sec! ● Bitsy is now comparable to Neo4J in read throughput* . ○ This is an apples-to-apples comparison since Neo4J is embedded and the graph is fully cached. ○ Most “bad” Neo4J benchmarks are taken when the graph doesn’t fit in memory. ○ Neo4J is extremely fast when the graph fits in memory -- and now, so is Bitsy!
- 10. Another read benchmark ● The following plot shows the traversal performance of Bitsy 1.5 vs Neo4J 1.9.2 in a multi-threaded setting on a bipartite graph with 1M vertices and out-degree of 3. ● Again, you can see that the performance is comparable.
- 11. Benchmarks for write ● As with 1.0 release, Bitsy’s write throughput is much higher than Neo4J because of the “No Seek” principle. ○ For more info, please refer to the project page at http://bitbucket.org/lambdazen/bitsy/
- 12. Wrap-up ● The 1.5 release introduces memory-efficient data structures and (mostly) lock-free reading to the Bitsy graph database. ○ With these improvements, Bitsy’s read performance is comparable to Neo4J’s cache. ○ Bitsy’s “No Seek” write algorithms continue to outperform other graph databases, including Neo4J. ● Bitsy is a dual-licensed product with ○ an AGPL license for open-source projects, and ○ a liberal unlimited-use OEM/end-user license for commercial projects. Details at lambdazen.com.