We are running 1.4.13 on wheezy. In the environment I am looking at there is positive correlation between gets and puts. The ration is something like 10 Gets : 15 Puts. The eviction spikes are also occurring at peak put times ( which kind of makes senses with the mem pressure ). I think the application is some kind of report generation tool - it's hard to say, my visibility into the team stuff is pretty low right now as I am a new hire.
On Sat, Aug 27, 2016 at 12:34 PM, dormando <[email protected]> wrote: > What version are you on and what're your startup options, out of > curiosity? > > A lot of the more recent features can help with memory efficiency, for > what it's worth. > > On Sat, 27 Aug 2016, Joseph Grasser wrote: > > > > > No problem, I'm trying cut down on cost. We're currently using a > dedicated model which works for us on a technical level but is expensive > (within budget but still expensive). > > > > We are experiencing weird spikes in evictions but I think that is the > result of developers abusing the service. > > > > Tbh I don't know what to make of the evictions yet. I'm gong to dig into > it on Monday though. > > > > > > On Aug 27, 2016 1:55 AM, "Ripduman Sohan" <[email protected]> > wrote: > > > > On Aug 27, 2016 1:46 AM, "dormando" <[email protected]> > wrote: > > > > > > Thank you for the tips guys! > > > > > > The limiting factor for us is actually memory > utilization. We are using the default configuration on sizable ec2 nodes > and pulling only > > like 20k qps per node. Which is fine > > > because we need to shard the key set over x servers > to handle the mem req (30G) per server. > > > > > > I should have looked into that before posting. > > > > > > I am really curious about network saturation though. > 200k gets at 1mb per get is a lot of traffic... how can you hit that mark > without > > saturation? > > > > Most people's keys are a lot smaller. In multiget > tests with 40 byte keys > > I can pull 20 million+ keys/sec out of the server. > probably less than > > 10gbps at that rate too. Tends to cap between 600k and > 800k/s if you need > > to do a full roundtrip per key fetch. limited by the > NIC. Lots of tuning > > required to get around that. > > > > > > I think (but may be wrong) the 200K TPS result is based on 1K values. > Dormando should be able to correct me. > > > > 20K TPS does seem a little low though. If you're bound by memory set > size have you thought of the cost/tradeoff benefits of using dedicated > servers for your memcache? > > I'm quite interested to find out more about what you're trying to > optimise. Is it minimising number of servers, maximising query rate, both, > none, etc? > > > > Feel free to reach out directly if you can't share this publicly. > > > > > > -- > > > > --- > > You received this message because you are subscribed to a topic in the > Google Groups "memcached" group. > > To unsubscribe from this topic, visit https://groups.google.com/d/ > topic/memcached/la-0fH1UzyA/unsubscribe. > > To unsubscribe from this group and all its topics, send an email to > [email protected]. > > For more options, visit https://groups.google.com/d/optout. > > > > -- > > > > --- > > You received this message because you are subscribed to the Google > Groups "memcached" group. > > To unsubscribe from this group and stop receiving emails from it, send > an email to [email protected]. > > For more options, visit https://groups.google.com/d/optout. > > > > > > -- > > --- > You received this message because you are subscribed to a topic in the > Google Groups "memcached" group. > To unsubscribe from this topic, visit https://groups.google.com/d/ > topic/memcached/la-0fH1UzyA/unsubscribe. > To unsubscribe from this group and all its topics, send an email to > [email protected]. > For more options, visit https://groups.google.com/d/optout. > -- --- You received this message because you are subscribed to the Google Groups "memcached" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
