Improve benchmark scripts

Summary:
This adds:
1) use of --level_compaction_dynamic_level_bytes=true
2) use of --bytes_per_sync=2M
The second is a big win for disks. The first helps in general.

This also adds a new test, fillseq with 32kb values to increase the peak
ingest and make it more likely that storage limits throughput.

Sample outpout from the first 3 tests - https://gist.github.com/mdcallag/e793bd3038e367b05d6f

Task ID: #

Blame Rev:

Test Plan:
Revert Plan:

Database Impact:

Memcache Impact:

Other Notes:

EImportant:

- begin *PUBLIC* platform impact section -
Bugzilla: #
- end platform impact -

Reviewers: igor

Reviewed By: igor

Subscribers: dhruba

Differential Revision: https://reviews.facebook.net/D37509

tools/benchmark.sh

@@ -46,6 +46,7 @@ duration=${DURATION:-0}
 num_keys=${NUM_KEYS:-$((1 * G))}
 key_size=20
 value_size=${VALUE_SIZE:-400}
+block_size=${BLOCK_SIZE:-4096}
 
 const_params="
   --db=$DB_DIR \
@@ -56,12 +57,14 @@ const_params="
   --num_levels=6 \
   --key_size=$key_size \
   --value_size=$value_size \
-  --block_size=4096 \
+  --block_size=$block_size \
   --cache_size=$cache_size \
   --cache_numshardbits=6 \
   --compression_type=zlib \
   --min_level_to_compress=3 \
   --compression_ratio=0.5 \
+  --level_compaction_dynamic_level_bytes=true \
+  --bytes_per_sync=$((2 * M)) \
   \
   --hard_rate_limit=3 \
   --rate_limit_delay_max_milliseconds=1000000 \
@@ -163,10 +166,10 @@ function run_fillseq {
        --threads=1 \
        --memtablerep=vector \
        --disable_wal=1 \
-       2>&1 | tee -a $output_dir/benchmark_fillseq.log"
+       2>&1 | tee -a $output_dir/benchmark_fillseq.v${value_size}.log"
-  echo $cmd | tee $output_dir/benchmark_fillseq.log
+  echo $cmd | tee $output_dir/benchmark_fillseq.v${value_size}.log
   eval $cmd
-  summarize_result $output_dir/benchmark_fillseq.log fillseq fillseq
+  summarize_result $output_dir/benchmark_fillseq.v${value_size}.log fillseq.v${value_size} fillseq
 }
 
 function run_change {

tools/run_flash_bench.sh

@@ -37,6 +37,7 @@
 #                   You can estimate the size of the test database from this,
 #                   NKEYS and the compression rate (--compression_ratio) set
 #                   in tools/benchmark.sh
+#   BLOCK_LENGTH  - value for db_bench --block_size
 #   CACHE_BYTES   - the size of the RocksDB block cache in bytes
 #   DATA_DIR      - directory in which to create database files
 #   LOG_DIR       - directory in which to create WAL files, may be the same
@@ -61,6 +62,7 @@ duration=${NSECONDS:-$((60 * 60))}
 nps=${RANGE_LIMIT:-10}
 vs=${VAL_SIZE:-400}
 cs=${CACHE_BYTES:-$(( 1 * G ))}
+bs=${BLOCK_LENGTH:-4096}
 
 # If no command line arguments then run for 24 threads.
 if [[ $# -eq 0 ]]; then
@@ -88,6 +90,7 @@ NUM_KEYS=$num_keys \
 DB_DIR=$db_dir \
 WAL_DIR=$wal_dir \
 VALUE_SIZE=$vs \
+BLOCK_SIZE=$bs \
 CACHE_SIZE=$cs"
 
 mkdir -p $output_dir
@@ -109,7 +112,12 @@ if [[ $do_setup != 0 ]]; then
   # Test 1: bulk load
   env $ARGS ./tools/benchmark.sh bulkload
 
-  # Test 2: sequential fill
+  # Test 2a: sequential fill with large values to get peak ingest
+  #          adjust NUM_KEYS given the use of larger values
+  env $ARGS BLOCK_SIZE=$((1 * M)) VALUE_SIZE=$((32 * K)) NUM_KEYS=$(( num_keys / 64 )) \
+       ./tools/benchmark.sh fillseq
+
+  # Test 2b: sequential fill with the configured value size
   env $ARGS ./tools/benchmark.sh fillseq
 
   # Test 3: single-threaded overwrite