Transcript Defense - Northwestern Networks Group
FAWN: A Fast Array of Wimpy Nodes
Presented by: Clint Sbisa & Irene Haque
Motivation Large-scale data-intensive applications Facebook, LinkedIn, Dynamo CPU-I/O Gap storage, network and memory bottlenecks low CPU utilization CPU Power slower CPUs execute more queries per second per Watt 1 billion vs. 100 million instructions per Joule inefficient energy saving techniques Memory Power
FAWN Data-intensive, computational simple workloads Small objects - 100B - 1KB Cluster of embedded CPUs using flash storage Efficient Fast random reads Slow random writes FAWN-KV Key-value storage Consistent Hashing FAWN-DS Data store Log structured
FAWN - DS Log-structure key-value store Contains all values in a key range for each virtual ID Maps 160-bit key Hash Index bucket =
i
low order index bits key fragment = next 15 low order bits 6 byte in-memory Hash Index stores frag and pointer
FAWN - DS Basic Functions: Store Lookup Delete Virtual Node Maintenance: Split Merge Compact Concurrent operations
FAWN - KV Consistent hashing of back-end VIDs Management node assigns each front-end to circular key space Front-end nodes manages its key space forwards out-of-range request Back-end nodes - VIDs contacts front-end when joining owns a key range
FAWN - KV Chain replication
FAWN - KV Join split key range pre-copy chain insertion log flush Leave merge key range Join into each chain
Individual Node Performance
• • Lookup speed Bulk store speed: 23.2 MB/s, or 96% of raw speed
Individual Node Performance
• • Put speed Compared to BerkeleyDB: 0.07 MB/s – shows necessity of log-based filesystems
Individual Node Performance
• Read- and write-intensive workloads
System Benchmarks
• System throughput and power consumption
Impact of Ring Membership Changes • Query throughput during node join and maintenance operations
Impact of Ring Membership Changes • Query latency
Alternative Architectures
• • • Large Dataset, Low Query → FAWN+Disk Small Dataset, High Query → FAWN+DRAM Middle Range → FAWN+SSD
Conclusion
• • Fast and energy efficient processing of random read intensive workloads Over an order of magnitude more queries per Joule than traditional disk-based systems