• Leonardo Marmol
• Swaminathan Sundararaman
• Nisha Talagala
• Raju Rangaswami

  Direct KV store

• Writes must have a latency of at least 20 ms
• Reads must have a latency of at most 55 ms

Using NVMKV own benchmarking tool we obtained the following results. This tools generates a randomized workload in which every operation is performed on a different key.

Raw data can be found here cache_data_nolocks.pdf

• Write Amplification

• YCSB Results

Modification to the workloads

• fieldcount = 1 (default 10)
• fieldlength = 1024 (default 100)
• recordcount = 1000000 (default 1000)
• operationcount = 1000000 (default 1000)
• Total Data Size = 976.56 MB

Core Properties: https://github.com/brianfrankcooper/YCSB/wiki/Core-Properties

Workloads Descriptions: https://github.com/brianfrankcooper/YCSB/wiki/Core-Workloads

Raw data can be found here: http://gladiator.cs.fiu.edu/ycsb_workloads/

LevelDB Options File: https://code.google.com/p/leveldb/source/browse/include/leveldb/options.h

Official information about LevelDB http://leveldb.googlecode.com/svn/trunk/doc/index.html

General information about LevelDB as well and tips and tricks about tuning http://docs.basho.com/riak/latest/ops/advanced/backends/leveldb/

YCSB/Mapkeeper: https://github.com/m1ch1/mapkeeper/tree/master/leveldb

According to these benchmarks ran with Mapkeeper, WiredTiger outperforms LevelDB in every test. Details here https://github.com/wiredtiger/wiredtiger/wiki/YCSB-Mapkeeper-benchmark

• When running YCSB, for every workload I get the following error message. This was solved by applying a patch.

  Error, atomic write failed: 22
  The output of ''dmesg'' shows the following 
  ...
  [962679.776274] <6>fioinf Fusion-io 825GB ioScale2 0000:05:00.0: Attach succeeded.
  [962680.107683] <6>fioinf Fusion-io 825GB ioScale2 0000:05:00.0: atomic write too large, nr_pages = 4224 max 1088)

• The cache is not printing stats at the end of a run. This was caused by the helper library not calling nvm_kv_close. It was fixed with the following change

  diff --git a/helper/src/main/fio_kv_helper.cpp b/helper/src/main/fio_kv_helper.cpp
  index 7cb7a5d..56ca52f 100644
  --- a/helper/src/main/fio_kv_helper.cpp
  +++ b/helper/src/main/fio_kv_helper.cpp
  @@ -127,6 +127,8 @@ void fio_kv_close(fio_kv_store_t *store)
          assert(store != NULL);
          assert(store->kv > 0);
   
  +       nvm_kv_close(store->kv);
  +
          if (store->fd) {
                  fsync(store->fd);
                  close(store->fd);

• 05/05/15: Final version comments discussion with Swami / Nisha
• 04/24/15: ATC Comments and plans for final version / shepherding
• 03/20/15: NVMKV bug & Future directions for NVM research
• 03/03/15: Parallelism bug with NVMKV
• 02/06/15: Parallelism bug with NVMKV
• 01/30/15: TBD
• 01/27/15: TBD
• 12/16/14: Next steps with ioDrive2 and paper writing
• 12/16/14: Next steps with ioDrive2 and paper writing
• 12/09/14: New machine data (needs more validation); plans for updating testbed with ioDrive2
• 12/02/14: New machine data (doesn't look right)
• 10/30/14: Writing tasks
• 10/14/14: Eval/writing next steps
• 10/07/14: Plans moving forward without Eurosys submission
• 10/03/14: Paper submission plans and writing tasks for Eurosys'15
• 09/17/14: Reviewing the current outline and paper writeup; next steps for writing
• 09/16/14: Outline/data discussion
• 09/10/14: Writing plans
• 09/05/14: Evaluation plans; writing plans
• 09/02/14: Plans for paper
• 08/28/14: Two papers instead of one?
• 08/19/14: Trade-offs detailed discussion and new plans for selling FAST paper
• 08/12/14: Selling FAST'15 paper as a trade-offs paper
• 08/05/14: FAST'15 paper writing: new direction to include memory/flash hybrid KV tradeoffs for cost/performance/energy
• 06/24/14: Feedback for FAST submission and plans
• 05/06/14: HotStorage'14 camera-ready final touches
• 04/29/14: HotStorage'14 camera-ready and FAST submission summer plan
• 04/15/14: HotStorage'14 reviewer feedback discussion
• 04/08/14: Initial multi-threading data for YCSB
• 04/01/14: FAST paper brainstorming
• 03/25/14: FAST paper – initial thoughts
• 03/11/14: HotStorage evaluation data and Writing
• 03/04/14: Plans for HotStorage submission
• 02/25/14: ext2/ext4 do *worse* with LevelDB; HotStorage submission discussion
• 02/11/14: Updates on level-db write amplification experiments
• 01/14/14: Time-series plots
• 01/06/14: Next steps in data collection
• 12/23/13: Analyzing new LevelDB and DirectKV-with-cache data; and plans for submission
• 11/15/13: Directkv / driver issues; Leveldb cache design discussion
• 11/04/13: Initial cache numbers
• 10/28/13: Updates and next steps for cache implementation
• 10/21/13: DirectKV vs. LevelDB evaluation configurations, implementing write-through caching in DirectKV
• 10/14/13: DirectKV vs. LevelDB (initial data), Alternative comparison setups, Importance of Write-back caching,