I also used RAM disk as well. And one thing I will try is to disable the ram cache of ATS - https://docs.trafficserver.apache.org/en/latest/admin-guide/storage/index.en.html#disabling-the-ram-cache
I think that will then make it a bit more straightforward to reason about the whole situation. Kit On Wed, Apr 7, 2021 at 10:04 AM Nick Dunkin <[email protected]> wrote: > Hi Rob, > > > > This is excellent information, thank you very much. > > > > I had considered a large RAM cache setting (ram_cache) with just a small > disk cache, but the ATS documentation was very vague about what this would > mean. The documentation refers to the ram cache as a place where popular > items are *promoted*, inferring (to me at least) that the size of the > disk cash would always have to be larger than the ram cache, in order make > full use of the ram cache. > > > > As you say though, I prefer the RAM disk path. > > > > Thanks again, > > > > Nick > > > > *From: *Robert O Butts <[email protected]> > *Reply-To: *"[email protected]" < > [email protected]> > *Date: *Wednesday, April 7, 2021 at 12:54 PM > *To: *"[email protected]" <[email protected]> > *Subject: *Re: RAM disks for cache? > > > > Yes, we have some similar servers in Production. We use ramdisks, it works > fine. You could also theoretically give ATS a large memory cache and small > disk (proxy.config.cache.ram_cache.size). But then that cache is shared > across all remaps, and you lose the control of different volumes for > different domains, and ATS just generally isn't really designed for that. > We've found ramdisks to work better in practice. > > I will note, our servers like this also have poor CPUs. We thought this > would be ok when we bought them, but it turns out you really kind of need > decent CPU for things like SSL, and poor CPUs also tend to have poor PCI > lanes, which are important. Our servers like this tend to cap out around > 10Gbps, despite +20Gbps NICs. I don't know if yours are similar, but if > your memory-poor servers are also CPU-poor, you might have the same > bottlenecks. And that's not ATS' fault, it seems unlikely any other caching > proxy could do better. > > > > > > On Wed, Apr 7, 2021 at 10:12 AM Nick Dunkin <[email protected]> > wrote: > > Hi, > > > > I am being asked if Traffic Server can be efficiently run (or run at all) > on a disk *poor*, memory *rich *server. The specifics are out of scope > for this forum, but the short version is that I’m being asked to make use > of some existing hardware. > > > > I was considering creating a RAM disk from the several hundred GB of > available memory on these servers. > > > > Does anyone have any experience of running Traffic Server this way? > > > > Thanks, > > > > Nick > > > > *Nick Dunkin* > > Director, Software Engineering > > Manager – Architecture and New Product Introduction > > *o: * *+1 678.258.4071* > > *e:* [email protected] > > > > [image: [email protected]] > >
