We tested up to the ulimit (~16K) of connections against a single server and performance was ok, but I would definitely try to do some serious load testing before I put a system into production that I knew was going to have that load from the get-go. The system degrades VERY ungracefully when you hit the ulimit for the process, so be sure to have enough ensemble nodes to spread those connections across that this won't happen. I think maybe there's a JIRA out to deal with this issue, not sure what the status is.
C -----Original Message----- From: Patrick Hunt [mailto:ph...@apache.org] Sent: Thursday, November 18, 2010 2:57 PM To: zookeeper-user@hadoop.apache.org Subject: Re: number of clients/watchers fyi: I haven't heard of anyone running over 10k sessions. I've tried 20k before and had issues, you may want to look at this sooner rather than later. * Server gc tuning will be an issue (be sure to use cms/incremental). * Be sure to disable clients accessing the leader (server configuration param). * You may need to use the Observers feature to scale out this large. Patrick On Thu, Nov 18, 2010 at 10:31 AM, Jeremy Hanna <jeremy.hanna1...@gmail.com> wrote: >>> Can you clarify what you mean when you say 10-100K watchers? Do you mean >>> 10-100K clients with 1 active watch, or some lesser number of clients with >>> more watches, or a few clients doing a lot of watches and other clients >>> doing other things? > > Probably 10-100K clients each with 1 or 2 active watches. The clients will > respond to watch events and sometimes initiate actions of their own. > >> here's a similar test setup I used: > > Thanks Patrick - it's really nice to have those numbers and test harness > basis. > > We're still in architecture mode so some of the details are still in flux, > but I think this gives us an idea. > > Thanks very much. > > On Nov 18, 2010, at 11:51 AM, Patrick Hunt wrote: > >> Camille, that's a very good question. Largest cluster I've heard about >> is 10k sessions. >> >> Jeremy - largest I've ever tested was a 3 server cluster with ~500 >> sessions. Each session created 10k znodes (100bytes each znode) and >> set 5 watches on each. So 5 million znodes and 25million watches. I >> then had the sessions delete the znodes and looked for the >> notifications. They were processed by the clients quite quickly (order >> of seconds) iirc. Note: this required some GC tuning on the servers to >> operate correctly (in particular cms and incremental gc was turned on >> and sufficient memory was allocated for the heaps). >> >> here's a similar test setup I used: >> http://wiki.apache.org/hadoop/ZooKeeper/ServiceLatencyOverview >> this is the latency tester tool >> https://github.com/phunt/zk-smoketest >> >> Patrick >> >> On Thu, Nov 18, 2010 at 9:44 AM, Fournier, Camille F. [Tech] >> <camille.fourn...@gs.com> wrote: >>> Can you clarify what you mean when you say 10-100K watchers? Do you mean >>> 10-100K clients with 1 active watch, or some lesser number of clients with >>> more watches, or a few clients doing a lot of watches and other clients >>> doing other things? >>> >>> -----Original Message----- >>> From: Jeremy Hanna [mailto:jeremy.hanna1...@gmail.com] >>> Sent: Thursday, November 18, 2010 12:15 PM >>> To: zookeeper-user@hadoop.apache.org >>> Subject: number of clients/watchers >>> >>> I had a question about number of clients against a zookeeper cluster. I >>> was looking at having between 10,000 and 100,000 (towards 100,000) watchers >>> within a single datacenter at a given time. Assuming that some fraction of >>> that number are active clients and the r/w ratio is well within the >>> zookeeper norms, is that number within the realm of possibility for >>> zookeeper? We're going to do testing and benchmarking and things, but I >>> didn't want to go down a rabbit hole if this is simply too much for a >>> single zookeeper cluster to handle. The numbers I've seen in blog posts >>> vary and I saw that the observers feature may be useful in this kind of >>> setting. >>> >>> Maybe I'm underestimating zookeeper or maybe I don't have enough >>> information to tell. I'm just trying to see if zookeeper is a good fit for >>> our use case. >>> >>> Thanks. >>> > >