FAWN: A Fast Array of Wimpy Nodes Presented by: Aditi Bose & Hyma Chilukuri
Motivation Large-scale data-intensive applications like high performance key-value storage systems are being used by Facebook, LinkedIn, Amazon with more regularity. Being I/O, Requiring RA over large DB, performing parallel, concurrent and mostly independent operations, requiring large clusters and storing small sized objects are several common features these workloads share. System performance: queries/sec Energy efficiency: queries/joule CPU performance and I/O bandwidth Gap : For data intensive computing workloads, storage, network and memory bandwidth bottlenecks lead to low CPU utilization Solution: wimpy processors to reduce I/O induced idle cycles CPU Power consumption: operating processors at higher freq requires more energy. techniques to mask CPU bottleneck cause energy inefficiency branch prediction, speculative execution – more processor die area Solution: slower CPUs execute more instructions per joule 1 billion vs. 100 million instructions per Joule
FAWN Efficient – 1W at heavy load Vs 10W at load Fast random reads – up to 175 times faster Slow random writes – updating a single page means erasing an entire block before writing the modified block in its place Cluster of embedded CPUs using flash storage Efficient – 1W at heavy load Vs 10W at load Fast random reads – up to 175 times faster Slow random writes – updating a single page means erasing an entire block before writing the modified block in its place FAWN-KeyValue nodes organized into a ring using consistent Hashing physical node is a collection of virtual node FAWN-DS Log structured key-value stores contains values for key range associated with VID
FAWN - DS Uses as in-memory Hash Index to map 160-bit key to a value stored in the data log stores only a fragment of the actual key. Hash Index bucket = i low order index bits key fragment = next 15 low order bits Each bucket -6 bytes - stores frag, valid bit and 4-byte pointer
FAWN - DS Basic Functions: Store Lookup Delete Concurrent operations Virtual Node Maintenance: Split Merge Compact
FAWN-KV organizes the back-end VIDs into a storage ring- structure using consistent hashing Management node assigns each front-end to circular key space Front-end node manages fraction of key-space manages the VID membership list forwards out-of-range request Back-end nodes – VIDs owns a key range contacts front-end when joining FAWN - KV
Chain replication FAWN - KV
Join split key range pre-copy chain insertion log flush Leave merge key range Join into each chain FAWN - KV
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
Alternative Architectures Large Dataset, Low Query → FAWN+Disk number of nodes dominated by storage capacity per node has the lowest total cost per GB Small Dataset, High Query → FAWN+DRAM number of nodes dominated by per node query capacity has the lowest cost for queries/sec Middle Range → FAWN+SSD best balance of storage capacity, query rate and total cost
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