False alarm. Looks like we had some debugging flags on that severely reduced performance. Sorry about that. Regards. -Jeremy
> On Jun 7, 2021, at 8:38 AM, Dave Marion <[email protected]> wrote: > > Jeremy, > > It seems that you have the ability to quickly run your test to determine if > a release is "good". Testing 1.8 ruled out a lot of commits for us to look > at. Would it be possible for you to test a few others so that we can try and > narrow it down even more. The following releases are after 1.8.0: > > 1.8.1 > 1.9.0 > 1.9.1 > 1.9.2 > 1.9.3 > 1.10.0 > 1.10.1 > > Could you test with the 1.9.2 release? > > >> On 06/05/2021 10:04 PM Kepner, Jeremy - LLSC - MITLL <[email protected]> >> wrote: >> >> >> I did a quick check with Accumulo 1.8 and I get the expected single node >> performance scalability. >> So between Accumulo 1.8 and 1.10.1 something changed that significantly >> slowed the performance. >> >>> On Jun 5, 2021, at 8:48 PM, Kepner, Jeremy - LLSC - MITLL >>> <[email protected]> wrote: >>> >>> Hi Dave, >>> I am looking into the Accumulo/Hadoop configuration. Hopefully it is as >>> simple as getting the settings the same. The hardware configurations is: >>> Dual Xeon Platinum 8260 2.4 GHz 48 cores, DDR4 2.93 GHz 192 GB RAM. I am >>> looking into the disk specs, but that shouldn't matter since the writes are >>> only a few megabytes. I also just tested on some older hardware that is >>> closer to what was used in the 2014 paper, and the single process ingest >>> rate is ~8x slower. >>> >>> Has anyone done any recent benchmarking of Accumulo 1.10+? >>> >>> Regards. -Jeremy >>> >>> >>>> On Jun 5, 2021, at 7:08 PM, Dave Marion <[email protected]> wrote: >>>> >>>> Jeremy, >>>> >>>> Are you able to share any details about the hardware and the Accumulo >>>> configuration? Is the Accumulo/Hadoop configuration the same as the prior >>>> test (no replication, WAL turned off, batch writer configuration, etc.) >>>> >>>> Dave >>>> >>>> On Sat, Jun 5, 2021 at 6:12 PM Kepner, Jeremy - LLSC - MITLL < >>>> [email protected]> wrote: >>>> >>>>> Has anyone benchmarked Accumulo 1.10.1? I have been looking into repeating >>>>> the measurements we did in 2014 with Accumulo 1.5 ( >>>>> https://arxiv.org/abs/1406.4923) using Accumulo 1.10.1 on a bigger system >>>>> with more modern hardware. Unfortunately, when I repeat the single node >>>>> measurements, there is no performance improvement from having multiple >>>>> ingestors inserting into different presplits of a table. I get 120K >>>>> inserts/sec with one ingestor and 2x60K inserts/sec with two ingestors. >>>>> In >>>>> 2014 we got linear speedup to ~6 ingestors, providing ~600K inserts/sec on >>>>> a single node. >>>>> >>>>> Regards. -Jeremy >>>
