[hyc_symas]

Microbenchmarks practically never map 1:1 to real world performance, since these libraries get embedded in much larger software systems that each have their own bottlenecks and inefficiencies. That should already be well understood when we call these things "microbenchmarks" and not "benchmarks". Meanwhile, compare all of the LevelDB/Kyoto/SQLite3 results we reported to the results the LevelDB authors originally reported - you'll find they are directly comparable. It may well be that read tests whose data sets are already fully cached in RAM are not representative of the underlying media speed. But (1) we're just trying to produce numbers using the same methodology as the original LevelDB benchmarks and (2) the full results show that even with all data fully cached in RAM, the type of underlying filesystem still had dramatic effects on the performance of each library.

We've done another bench recently (not yet prettied up for publication) with a VM with 32GB of RAM and 4 CPU cores. On an LDAP database with 5 million entries, the full DB occupies 9GB of RAM and this VM can return 33,000 random queries/second ad nauseum. Scaling the DB up to 25 million entries, the DB requires 40GB of RAM and the random query rate across all 25M drops down to around 20,000/sec. Tests like these are more about the price/quality of your hard drive than the quality of your software. As publishable results they're much less referenceable since any particular site's results will have much greater variance, and the only thing for a conscientious party to do is measure it on their own system anyway.

[jules]

Fair enough, I don't see any problem with in memory benchmarks, as long as they are marked as such, and if you're comparing apples to apples. The best way to do this being actually using the data from the read queries to do some trivial operation like computing the XOR hash -- that would still be a best case for your library yet still real world.

I've read the papers on your DB and they are quite interesting. What do you think about the work in the paper linked in this post? It's unfortunate that they just compare with skip lists. I don't think anybody seriously believes skip lists are a good idea ever, so it's a bit of a straw man at this point (though I may be wrong).

[hyc_symas]

I find the Bw-tree work pretty underwhelming. It's not a transactional engine; the fact that it can operate "latch-free" is irrelevant since it must be wrapped by some higher concurrency management layer to truly be useful. The delta-update mechanism is probably more write-efficient than LMDB, which could be a point in their favor. The fact that they still rely on an application-level page cache manager is a strong negative - you can never manage the cache more accurately and more efficiently than the kernel, which gets hardware assist for free. Overall, it's an awful lot of noise being made over not much actual innovation.