In a write heavy workload, where L0-L1 compaction is the bottleneck, the level_compaction_dynamic_level_bytes flag can cause serious stalls caused by huge (40 Gb) L0-L1 compactions.
Can be seen in the pic below:
This problem exists mainly with the speedb delay mechanism (controlled by the flag use_dynamic_delay).
Further research needs to be done in order to understand and fix this behaviour.
Until then, Turn off the level_compaction_dynamic_level_bytes flag by default.
The cmd to reproduce the above Fig. is:
db_bench --compression_type=None -db=/data/ -num=40000000 -value_size=1000 -key_size=16 -report_interval_seconds=1 -stats_dump_period_sec=120 -num_column_families=1 -write_buffer_size=67108864 -histogram --delayed_write_rate=536870912 -max_write_buffer_number=4 -db_write_buffer_size=0 -max_background_compactions=8 -cache_size=8388608 -max_background_flushes=4 -bloom_bits=10 -benchmark_read_rate_limit=0 -benchmark_write_rate_limit=0 -report_file=fillrandom.csv --disable_wal=true --benchmarks=fillrandom,levelstats,memstats --memtablerep=skip_list -compaction_readahead_size=0 ' and then overwrite:
db_bench --compression_type=None -db=/data/ -num=40000000 -value_size=1000 -key_size=16 -report_interval_seconds=1 -stats_dump_period_sec=120 -num_column_families=1 -write_buffer_size=67108864 -histogram --delayed_write_rate=536870912 -max_write_buffer_number=4 -db_write_buffer_size=0 -max_background_compactions=8 -cache_size=8388608 -max_background_flushes=4 -bloom_bits=10 -duration=900 --use_existing_db -threads=50 -duration=900 -benchmark_read_rate_limit=0 -benchmark_write_rate_limit=0 -report_file=overwrite.csv --benchmarks=overwrite,levelstats,memstats --memtablerep=skip_list -compaction_readahead_size=0
and compare tolevel_compaction_dynamic_level_bytes = 0`
In a write heavy workload, where L0-L1 compaction is the bottleneck, the![image](https://github.com/speedb-io/speedb/assets/105581454/1d11aa7d-3e6b-4b2e-ba07-0c70689c1a4b)
level_compaction_dynamic_level_bytes
flag can cause serious stalls caused by huge (40 Gb) L0-L1 compactions. Can be seen in the pic below:This problem exists mainly with the speedb delay mechanism (controlled by the flag use_dynamic_delay). Further research needs to be done in order to understand and fix this behaviour.
Until then, Turn off the level_compaction_dynamic_level_bytes flag by default.
The cmd to reproduce the above Fig. is:
db_bench --compression_type=None -db=/data/ -num=40000000 -value_size=1000 -key_size=16 -report_interval_seconds=1 -stats_dump_period_sec=120 -num_column_families=1 -write_buffer_size=67108864 -histogram --delayed_write_rate=536870912 -max_write_buffer_number=4 -db_write_buffer_size=0 -max_background_compactions=8 -cache_size=8388608 -max_background_flushes=4 -bloom_bits=10 -benchmark_read_rate_limit=0 -benchmark_write_rate_limit=0 -report_file=fillrandom.csv --disable_wal=true --benchmarks=fillrandom,levelstats,memstats --memtablerep=skip_list -compaction_readahead_size=0 ' and then overwrite:
db_bench --compression_type=None -db=/data/ -num=40000000 -value_size=1000 -key_size=16 -report_interval_seconds=1 -stats_dump_period_sec=120 -num_column_families=1 -write_buffer_size=67108864 -histogram --delayed_write_rate=536870912 -max_write_buffer_number=4 -db_write_buffer_size=0 -max_background_compactions=8 -cache_size=8388608 -max_background_flushes=4 -bloom_bits=10 -duration=900 --use_existing_db -threads=50 -duration=900 -benchmark_read_rate_limit=0 -benchmark_write_rate_limit=0 -report_file=overwrite.csv --benchmarks=overwrite,levelstats,memstats --memtablerep=skip_list -compaction_readahead_size=0and compare to
level_compaction_dynamic_level_bytes = 0`