- YCSB - real benchmarking suit for real workloads, because both uniform and skewed are synthetic
- Fix the skewed distribution so we can take a look at the real numbers (maybe even enable partitioned mode under some conditions)
- Clean up the code to plug in different kv sizes and algorithms
-
We need to dynamically create consumer (delegation) threads, if the proportion of writes goes up we need to create more [Daman] -- otherwise we suck on growing number of writes
-
We need to re-balance the partitions to distribute the load proportionally [Vikram] --- otherwise we suck on skew
-- Mapping of partitions to threads and change it dynamically [comes from consistent hashing]
- Add bandwidth to the plots
- Hash join (run to see the results) [Niv]
-- We should add some real tables along with <8B, 8B>, e.g., <256B, 1024B>
- Aggregation (add aggregation on 20 byte keys -- matches genomic workloads)
- KV store over the network via DPDK (Mica-like kv-store), compare agains Mica?