News for hackers. Stuff that matters.
Re:LMDB: Intel Optane SSD Microbenchmark by wtallis on Wednesday November 07, 2018 @11:46PM (#18404385)
Write latency can be lower on flash-based SSDs than Optane because the former tend to have DRAM caches, but Optane SSDs have little to no caching. The difference isn't enough to produce that large of a discrepancy, so I suspect that the data being written to the Optane SSD was either not being batched into blocks of the same size as for the flash SSDs, or that fewer operations were being queued simultaneously for the Optane SSDs. I wonder if there were some mkfs or mount options that defaulted to a different value for the Optane SSD. I'm also curious whether the NVMe driver was used in polling mode.
Re:LMDB: Intel Optane SSD Microbenchmark by hyc_symas on Wednesday November 07, 2018 @11:51PM (#18404394)
Since the loader was performing asynchronous writes, actual batching would be whatever the OS is doing when it eventually flushes dirty pages to the underlying device. I can't imagine the OS using different queue depths during the RocksDB run than the LMDB run. mkfs/mount options were identical for both. We didn't touch the NVMe driver at all, so whatever its default behavior is, was used on all tests.

By the way - I think you can still get free access to run your own tests. https://twitter.com/IntelStorage/status/1010284314121129985

Let me know if you want to try to rerun these, and doublecheck the results. Also, we found that 4.x kernels performed significantly better than 3.x kernels on these tests.

Re:LMDB: Intel Optane SSD Microbenchmark by wtallis on Thursday November 08, 2018 @12:09AM (#18404450)
I already have the hardware needed to run my own tests (albeit only one P4800X), but I've been almost exclusively focusing on consumer SSDs recently. I'll have some time later this month to try to replicate your results. I don't have experience setting up RocksDB or LMDB. Is the process of getting your benchmark running pretty self-explanatory?
Re:LMDB: Intel Optane SSD Microbenchmark by hyc_symas on Thursday November 08, 2018 @12:16AM (#18404473)
All the shell scripts I used are in the data.tgz file linked at the bottom of the page. You should be able to run after editing a few pathnames. Compiling the dbbench drivers should be simple enough; I may be able to send you binaries if not.