search

the-data-lab / GraphOne

"GraphOne: A Data Store for Real-time Analytics on Evolving Graphs", Usenix FAST'19
GNU General Public License v3.0
58 stars 31 forks source link

Performance regression #17

Open whatsthecraic opened 4 years ago

whatsthecraic commented 4 years ago

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 623d3f88099f9244e9c370e97c4afbe40d2ffb6c 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

pradeep-k commented 4 years ago

Hi Dean, I am not sure why performance regression is failing. I can take a look into it by month end only.

If you have any functionality bug, let me know. Otherwise I will check the performance issue by month end.

whatsthecraic commented 4 years ago

Hey there,

thank you for the quick reply. It doesn't affect any functionality. I'll wait for a future update then.

Thanks, Dean