airmail icon indicating copy to clipboard operation
airmail copied to clipboard

Performance on high end server - RPS

Open do-me opened this issue 9 months ago • 4 comments

I know that the scope of airmail is to run a geocoder on cheap hardware with acceptable latency but I was wondering if you had a high end server available, what kind of performance could you get out of a setup where the index lives in the system's RAM? Did you ever run any tests or do you remember out of the top of your head how performant it might be?

My aim is to get a many RPS out of common open source geocoders as possible and then refining the results.

E.g. with Photon on an AMD server with 16 cores and 128Gb RAM I get around 250 RPS for batch requests which is already pretty good.

If airmail had much higher throughput, I might be able to allocate some of my time into some of the problems with German (or maybe European?) address stemmers e.g. https://github.com/ellenhp/airmail/issues/9#issuecomment-2651486380

do-me avatar Apr 11 '25 15:04 do-me

I haven't run any tests but I should be able to do that side by side with photon at least (as always, thank you to graphhopper for publishing photon indexes!). Pelias is more involved to build an index for and I'm not sure I have bandwidth for that right now.

ellenhp avatar Apr 11 '25 19:04 ellenhp

For what it's worth, I have been running the Airmail demo with local-ish storage (cephfs on spinning rust) for a while now. We'll see how it does on NVMe. I think there's a good chance it'll be competitive, not sure about better though.

ellenhp avatar Apr 11 '25 19:04 ellenhp

Really curious about the results! When I find the time, I'll also test locally on my SSD and share the RPS.

do-me avatar Apr 11 '25 20:04 do-me

@do-me I'm interested as well

Because with the following hw settings, I'm having issues in deploying it :

https://github.com/ellenhp/airmail/issues/34#issuecomment-2838773710

raphael10-collab avatar May 20 '25 09:05 raphael10-collab