Hi I've noticed the decreased performance in 4MB experiment as well. For Jerasure for example, my profile is a simple k=4 m=5 configuration. Being more precise, the decrease occurs near the end of the reads, when ~80% of reads were done. Looks like the amount of reads reduces toward the end of the experiment, and it create a distortion of the avg MB/s and avg lat. First, when reaching 80% of total reads, it is visible that not every sec has a read done, meaning we done have a real value per each sec. Second, when there is no sample, MB/s is regarded as 0, and this value is inserted into avg MB/s - This is clearly incorrect! My questions is this: What happens toward the end of the experiment (80%) causing the reads to reduce drastically and the latency to increase? Thanks, Oleg On Thu, Aug 10, 2017 at 7:32 AM, Kyle Bader <kyle.bader@xxxxxxxxx> wrote: > 64MB is pretty large for a rados bench object - most of the interfaces > above rados stripe at a 4MB boundary. It could be that larger sized > writes like those require adjusting various knobs in the OSD. > > Do you see the same thing happen with 4MB work workloads? > > What are your EC profile(s)? > > On Wed, Aug 9, 2017 at 4:44 AM, Oleg Kolosov <olekol@xxxxxxxxx> wrote: >> Hi >> I've began experimenting with rados bench and I've noticed bizarre >> behavior for jerasure and lrc plugins. I'm writing very large amounts >> of data: ~2000 objects of size 64MB. >> When reading the data with 9 parallel client, I'm noticing a linear >> increase in latency around the middle of the read experiment (latency >> was steady up to that point). >> Is that an expected behavior? >> How can it be explained? >> >> Thanks, >> Oleg >> -- >> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in >> the body of a message to majordomo@xxxxxxxxxxxxxxx >> More majordomo info at http://vger.kernel.org/majordomo-info.html > > > > -- > > Kyle Bader -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
