Just to clarify. Yes, this is a constant load via a test application. The test scenario is live segmented video content (Apple HLS), so we have small, plain text, manifest files and large(ish) video files (3MB). There is also constant cache churn because the content is live video. We have set RAM cache and disk cache deliberately small (1GB each) to observe behavior as we churn cache.
The graph I attached was for ioBufAllocator[0], but we do see similar trends on [4] and [5]. We do see expected behavior, i.e. plateauing, from the other ioBufAllocators. Sorry for not making that clear in my initial email. Thanks, Nick From: Alan Carroll <[email protected]> Reply-To: "[email protected]" <[email protected]> Date: Wednesday, May 24, 2017 at 11:39 AM To: "[email protected]" <[email protected]> Subject: Re: Understanding ioBufAllocator behvaiour That can certainly happen and is a known problem, but it looked like Nick's scenario was a constant load via a test application and he saw unbounded growth in a single iobuf bucket. For threads, there is a single global pool and each thread keeps a smaller pool from the global one (via the ProxyAllocator instances). The ProxyAllocator has a high and low water mark - when the # of items in the thread exceeds the high water mark they are released back to the global pool until there are only low water mark items left. The values for these are in the 128-512 range, so not on the same scale as this memory growth. There's been lots of discussion about jemalloc. What we lack is production performance data to see what the impact would be. We're working on that. As far as I understand it (Phil and Leif know more) we would keep the ProxyAllocators but instead of releasing to a global pool the memory would be released to jemalloc for re-use, thereby strongly bounding the amount of memory in a particular iobuf bucket. On Wednesday, May 24, 2017, 10:29:32 AM CDT, Kapil Sharma (kapsharm) <[email protected]> wrote: On plateauing - not necessarily; we do see the memory consumption increasing continuously in our deployments as well. It depends on the pattern of segment sizes over time. ATS uses power of 2 allocators for memory pool - there are 15 of those, ranging from 128bytes to 2M if my memory serves me right - and these are per thread! ATS will choose an optimal allocator for the segments. As Alan mentioned, once chunk are allocated, they are never freed. Here is a totally artificial example just to make the point (please correct if my understanding is flawed): * the traffic pattern was such that initially only 2M allocators were used then ATS will keep allocating 2M chunks until RAM cache limit (lets say it is 64GB) is reached. * Now traffic pattern changed (smaller fragment requests), and only 1M allocators are used, ATS will now keep allocating 1M chunks, again capping at 64GB. But in the end ATS would have allocated 128GB well over RAM cache size limit…. In the past a there was some prototype of reclaimable buffer support added in ATS, but I believe it was removed in 7.0? Also there is recent discussion of adding jmalloc? On May 24, 2017, at 11:01 AM, Alan Carroll <[email protected]<mailto:[email protected]>> wrote: One issue is that memory never moves between the iobuf sizes. Once a chunk of memory is used for a specific iobuf slot, it's there forever. But unless something is leaking, the total size should eventually plateau, certainly within less than a day if you have a basically constant load. There will be some growth due to blocks being kept in thread local allocation pools, but again that should level in less time than you've run. On Wednesday, May 24, 2017, 9:50:39 AM CDT, Dunkin, Nick <[email protected]<mailto:[email protected]>> wrote: Hi Alan, This is 7.0.0 I only see this behavior on ioBufAllocator[0], [4] and [5]. The other ioBufAllocators’ usage looks as I would expect (i.e. allocated goes up then flat), so I was thinking it was more likely something to do with my configuration or use-case. I’d also just like to understand, at a high level, how the ioBufAllocators are used. Thanks, Nick From: Alan Carroll <[email protected]<mailto:[email protected]>> Reply-To: "[email protected]<mailto:[email protected]>" <[email protected]<mailto:[email protected]>> Date: Wednesday, May 24, 2017 at 10:33 AM To: "[email protected]<mailto:[email protected]>" <[email protected]<mailto:[email protected]>> Subject: Re: Understanding ioBufAllocator behvaiour Honestly it sounds like a leak. Can you specify which version of Traffic Server this is? On Wednesday, May 24, 2017, 8:22:46 AM CDT, Dunkin, Nick <[email protected]<mailto:[email protected]>> wrote: Hi I have a load test that I’ve been running for a number of days now. I’m using the memory dump logging in traffic.out and I’m trying to understand how Traffic Server allocates and reuses memory. I’m still quite new to Traffic Server. Nearly all of the memory traces look as I would expect, i.e. memory is allocated and reused over the lifetime of the test. However my readings from ioBufAllocator[0] show a continual increase in allocated AND used. I am attaching a graph. (FYI – This graph covers approximately 3 days of continual load test.) I would have expected to start seeing reuse in ioBufAllocator by now, like I do in the other ioBufAllocators. Can someone help me understand what I’m seeing? Many thanks, Nick Dunkin Nick Dunkin Principal Engineer o: 678.258.4071 e: [email protected]<mailto:[email protected]> 4375 River Green Pkwy # 100, Duluth, GA 30096, USA <image001.png> <image001.png>
