Why does the DiffKV outperform than PebblesDB for range query performance?

[zjs1224522500]

After reading the paper, I found that DiifKV use the vTree structure, similar to PebblesDB, to organize the middle-size key-value pairs. For range query, compared with PebblesDB, DiffKV seems to have no obvious advantage, except for the scan-optimized merge.

Why does the DiffKV outperform than PebblesDB for range query performance? Is this due to the scan-optimized merge strategy? Or is it caused by the mixed size workload? For middle-size key-value pairs, does the performance advantage of DiffKV range queries still exist?

[ysjyx7]

I think the difference between the scan performance of Pebblesdb and Diffkv in the evaluation results in section 5 of this paper seems to be small. Which result shows that DiffKV outperform than PebblesDB for range query performance?

[zjs1224522500]

The difference is small in range query of microbenchmarks, However, it seems to have a larger performance gap for YCSB-E, even though YCSB-E contains a small number of write operations.

[ysjyx7]

In ./bench_tools/YCSB-C/configDir/pebblesdb_config.ini, I find that the seekCompaction was set to false. Maybe this configuration will cause pebblesdb to have a guard with many sstables when there are some write operations in workload? Maybe this will reduce the scan performance of PebblesDB in YCSB-E. However, DiffKV can still use Scan Trigger Merge to improve data order to get better scan performance.