Performance regression #17
Description
Hi there,
as I've repeated another round of our internal experiments, I noticed a regression compared to the results I obtained back in February. For instance I am accounting a difference of ~2.9x for insertions on graph500 with s.f. 24 on a relatively modern dual socket Xeon, with 20 cores in total. My current branch tracks 623d3f8 after issue #16, while the results from February are likely based on the master branch that was available at the end of December'19, after issue #11 , but unfortunately I didn`t record the precise commit that was used at the time.
The throughput obtained is the same as before with up to two threads to execute the insertions. After that, the scalability doesn't improve any more in my current measurements, while in February I obtained the best results with 12 threads.
I wonder what is the reason for this difference and if there is something we can do to improve the current state. The patch for #16 only alters the creation of static views, which are not even used in this experiment.
Thanks,
Dean