Yes, it seems Hadoop framework did not consume all offered resources: if framework launch task (1 CPUs) on offer (10 CPUs), the other 9 CPUs will return back to master (recoverResouces).
---- Da (Klaus), Ma (马达) | PMP® | Advisory Software Engineer Platform OpenSource Technology, STG, IBM GCG +86-10-8245 4084 | klaus1982...@gmail.com | http://k82.me On Thu, Jan 21, 2016 at 6:46 PM, Tom Arnfeld <t...@duedil.com> wrote: > Thanks everyone! > > Stephan - There's a couple of useful points there, will definitely give it > a read. > > Klaus - Thanks, we're running a bunch of different frameworks, in that > list there's Hadoop MRv1, Apache Spark, Marathon and a couple of home grown > frameworks we have. In this particular case the Hadoop framework is the > major concern, as it's designed to continually accept offers until it has > enough slots it needs. With the example I gave above, we observe that the > master is never sending any sizeable offers to some of these frameworks > (the ones with the larger shares), which is where my confusion stems from. > > I've attached a snippet of our active master logs which show the activity > for a single slave (which has no active executors). We can see that it's > cycling though sending and recovering declined offers from a selection of > different frameworks (in order) but I can say that not all of the > frameworks are receiving these offers, in this case that's the Hadoop > framework. > > > On 21 January 2016 at 00:26, Klaus Ma <klaus1982...@gmail.com> wrote: > >> Hi Tom, >> >> Which framework are you using, e.g. Swarm, Marathon or something else? >> and which language package are you using? >> >> DRF will sort role/framework by allocation ratio, and offer all >> "available" resources by slave; but if the resources it too small (< >> 0.1CPU) or the resources was reject/declined by framework, the resources >> will not offer it until filter timeout. For example, in Swarm 1.0, the >> default filter timeout 5s (because of go scheduler API); so here is case >> that may impact the utilisation: the Swarm got one slave with 16 CPUS, but >> only launch one container with 1 CPUS; the other 15 CPUS will return back >> to master and did not re-offer until filter timeout (5s). >> I had pull a request to make Swarm's parameters configurable, refer to >> https://github.com/docker/swarm/pull/1585. I think you can check this >> case by master log. >> >> If any comments, please let me know. >> >> ---- >> Da (Klaus), Ma (马达) | PMP® | Advisory Software Engineer >> Platform OpenSource Technology, STG, IBM GCG >> +86-10-8245 4084 | klaus1982...@gmail.com | http://k82.me >> >> On Thu, Jan 21, 2016 at 2:19 AM, Tom Arnfeld <t...@duedil.com> wrote: >> >>> Hey, >>> >>> I've noticed some interesting behaviour recently when we have lots of >>> different frameworks connected to our Mesos cluster at once, all using a >>> variety of different shares. Some of the frameworks don't get offered more >>> resources (for long periods of time, hours even) leaving the cluster under >>> utilised. >>> >>> Here's an example state where we see this happen.. >>> >>> Framework 1 - 13% (user A) >>> Framework 2 - 22% (user B) >>> Framework 3 - 4% (user C) >>> Framework 4 - 0.5% (user C) >>> Framework 5 - 1% (user C) >>> Framework 6 - 1% (user C) >>> Framework 7 - 1% (user C) >>> Framework 8 - 0.8% (user C) >>> Framework 9 - 11% (user D) >>> Framework 10 - 7% (user C) >>> Framework 11 - 1% (user C) >>> Framework 12 - 1% (user C) >>> Framework 13 - 6% (user E) >>> >>> In this example, there's another ~30% of the cluster that is >>> unallocated, and it stays like this for a significant amount of time until >>> something changes, perhaps another user joins and allocates the rest.... >>> chunks of this spare resource is offered to some of the frameworks, but not >>> all of them. >>> >>> I had always assumed that when lots of frameworks were involved, >>> eventually the frameworks that would keep accepting resources indefinitely >>> would consume the remaining resource, as every other framework had rejected >>> the offers. >>> >>> Could someone elaborate a little on how the DRF allocator / sorter >>> handles this situation, is this likely to be related to the different users >>> being used? Is there a way to mitigate this? >>> >>> We're running version 0.23.1. >>> >>> Cheers, >>> >>> Tom. >>> >> >> >