Re: Some set() calls taking way too long
Few things that may help understanding your problem: 1. What is the status of your slabs allocation, is there enough room to all slabes ? 2. Do you see increase in the requests rate when your Memcached memory is becoming full with objects ? 3. How many threads are configured ? On Tue, Jul 17, 2012 at 1:11 PM, David Morel david.mo...@amakuru.netwrote: hi memcached users/devvers, I'm seeing occasional slowdowns (tens of milliseconds) in setting some keys on some big servers (80GB RAM allocated to memcached) which contain a large number of keys (many millions). The current version I use is 1.4.6 on RH6. The thing is once I bounce the service (restart, not flush_all), everything becomes fine again. So could a large number of keys be the source of the issue (some memory allocation slowdown or something)? I don't see that many evictions on the box, and anyway, evicting an object to make room for another shouldn't take long, should it? Is there a remote possibility the large number of keys is at fault and splitting the daemons, like 2 or more instances per box, would fix it? Or is that a known issue fixed in a later release? Thanks for any insight. David Morel -- Yiftach Shoolman +972-54-7634621
Re: Some set() calls taking way too long
On Tuesday, July 17, 2012 12:26:16 PM UTC+2, Yiftach wrote: Few things that may help understanding your problem: 1. What is the status of your slabs allocation, is there enough room to all slabes ? This happens when the memory gets close to full. however there is not a large number of evictions. I would expect evictions to be made whenever needed, but not the process of making room for 1 object to take half a second. 2. Do you see increase in the requests rate when your Memcached memory is becoming full with objects ? I don't think so, why would that be the case, it's application dependent, not server, right? 3. How many threads are configured ? the default 4 On Tue, Jul 17, 2012 at 1:11 PM, David Morel david.mo...@amakuru.netwrote: hi memcached users/devvers, I'm seeing occasional slowdowns (tens of milliseconds) in setting some keys on some big servers (80GB RAM allocated to memcached) which contain a large number of keys (many millions). The current version I use is 1.4.6 on RH6. The thing is once I bounce the service (restart, not flush_all), everything becomes fine again. So could a large number of keys be the source of the issue (some memory allocation slowdown or something)? I don't see that many evictions on the box, and anyway, evicting an object to make room for another shouldn't take long, should it? Is there a remote possibility the large number of keys is at fault and splitting the daemons, like 2 or more instances per box, would fix it? Or is that a known issue fixed in a later release? Thanks for any insight. David Morel -- Yiftach Shoolman +972-54-7634621
Re: Some set() calls taking way too long
Re #2 - when more objects are stored in the cache, the hit ratio should be higher-- the application might run faster, i.e. with less DB accesses. Anyway, it sounds to me more related to slabs allocation, as only reset solves it (not flush) if I understood u well. Does it happen on any object size or on specific object size range ? On Tue, Jul 17, 2012 at 4:04 PM, David Morel david.mo...@amakuru.netwrote: On Tuesday, July 17, 2012 12:26:16 PM UTC+2, Yiftach wrote: Few things that may help understanding your problem: 1. What is the status of your slabs allocation, is there enough room to all slabes ? This happens when the memory gets close to full. however there is not a large number of evictions. I would expect evictions to be made whenever needed, but not the process of making room for 1 object to take half a second. 2. Do you see increase in the requests rate when your Memcached memory is becoming full with objects ? I don't think so, why would that be the case, it's application dependent, not server, right? 3. How many threads are configured ? the default 4 On Tue, Jul 17, 2012 at 1:11 PM, David Morel david.mo...@amakuru.netwrote: hi memcached users/devvers, I'm seeing occasional slowdowns (tens of milliseconds) in setting some keys on some big servers (80GB RAM allocated to memcached) which contain a large number of keys (many millions). The current version I use is 1.4.6 on RH6. The thing is once I bounce the service (restart, not flush_all), everything becomes fine again. So could a large number of keys be the source of the issue (some memory allocation slowdown or something)? I don't see that many evictions on the box, and anyway, evicting an object to make room for another shouldn't take long, should it? Is there a remote possibility the large number of keys is at fault and splitting the daemons, like 2 or more instances per box, would fix it? Or is that a known issue fixed in a later release? Thanks for any insight. David Morel -- Yiftach Shoolman +972-54-7634621 -- Yiftach Shoolman +972-54-7634621
Re: Some set() calls taking way too long
On 17 Jul 2012, at 15:33, Yiftach Shoolman wrote: Re #2 - when more objects are stored in the cache, the hit ratio should be higher-- the application might run faster, i.e. with less DB accesses. no relation here, I was really mentioning bare set() calls timing, regardless of anything else. Anyway, it sounds to me more related to slabs allocation, as only reset solves it (not flush) if I understood u well. all the slabs are already allocated, with a moderate rate of evictions. so it's not slab allocation either, it's just one key now and then, regardless (it seems) of the key itself or the size of the object. Does it happen on any object size or on specific object size range ? anything, really. puzzling, hey? thanks! On Tue, Jul 17, 2012 at 4:04 PM, David Morel david.mo...@amakuru.netwrote: On Tuesday, July 17, 2012 12:26:16 PM UTC+2, Yiftach wrote: Few things that may help understanding your problem: 1. What is the status of your slabs allocation, is there enough room to all slabes ? This happens when the memory gets close to full. however there is not a large number of evictions. I would expect evictions to be made whenever needed, but not the process of making room for 1 object to take half a second. 2. Do you see increase in the requests rate when your Memcached memory is becoming full with objects ? I don't think so, why would that be the case, it's application dependent, not server, right? 3. How many threads are configured ? the default 4 On Tue, Jul 17, 2012 at 1:11 PM, David Morel david.mo...@amakuru.netwrote: hi memcached users/devvers, I'm seeing occasional slowdowns (tens of milliseconds) in setting some keys on some big servers (80GB RAM allocated to memcached) which contain a large number of keys (many millions). The current version I use is 1.4.6 on RH6. The thing is once I bounce the service (restart, not flush_all), everything becomes fine again. So could a large number of keys be the source of the issue (some memory allocation slowdown or something)? I don't see that many evictions on the box, and anyway, evicting an object to make room for another shouldn't take long, should it? Is there a remote possibility the large number of keys is at fault and splitting the daemons, like 2 or more instances per box, would fix it? Or is that a known issue fixed in a later release? Thanks for any insight. David Morel -- Yiftach Shoolman +972-54-7634621 -- Yiftach Shoolman +972-54-7634621 David Morel -- Booking.com http://www.booking.com/ Lyon office: 93 rue de la Villette, 5e étage, F-69003 Lyon phone:+33 4 20 10 26 63 gsm:+33 6 80 38 56 83
Re: Some set() calls taking way too long
One more guess - could it be that you are limited by the network bandwidth of your server ? I came across many Memcached deployments that this was the case On Tue, Jul 17, 2012 at 7:47 PM, David Morel david.mo...@amakuru.netwrote: On 17 Jul 2012, at 15:33, Yiftach Shoolman wrote: Re #2 - when more objects are stored in the cache, the hit ratio should be higher-- the application might run faster, i.e. with less DB accesses. no relation here, I was really mentioning bare set() calls timing, regardless of anything else. Anyway, it sounds to me more related to slabs allocation, as only reset solves it (not flush) if I understood u well. all the slabs are already allocated, with a moderate rate of evictions. so it's not slab allocation either, it's just one key now and then, regardless (it seems) of the key itself or the size of the object. Does it happen on any object size or on specific object size range ? anything, really. puzzling, hey? thanks! On Tue, Jul 17, 2012 at 4:04 PM, David Morel david.mo...@amakuru.net** wrote: On Tuesday, July 17, 2012 12:26:16 PM UTC+2, Yiftach wrote: Few things that may help understanding your problem: 1. What is the status of your slabs allocation, is there enough room to all slabes ? This happens when the memory gets close to full. however there is not a large number of evictions. I would expect evictions to be made whenever needed, but not the process of making room for 1 object to take half a second. 2. Do you see increase in the requests rate when your Memcached memory is becoming full with objects ? I don't think so, why would that be the case, it's application dependent, not server, right? 3. How many threads are configured ? the default 4 On Tue, Jul 17, 2012 at 1:11 PM, David Morel david.mo...@amakuru.net* *wrote: hi memcached users/devvers, I'm seeing occasional slowdowns (tens of milliseconds) in setting some keys on some big servers (80GB RAM allocated to memcached) which contain a large number of keys (many millions). The current version I use is 1.4.6 on RH6. The thing is once I bounce the service (restart, not flush_all), everything becomes fine again. So could a large number of keys be the source of the issue (some memory allocation slowdown or something)? I don't see that many evictions on the box, and anyway, evicting an object to make room for another shouldn't take long, should it? Is there a remote possibility the large number of keys is at fault and splitting the daemons, like 2 or more instances per box, would fix it? Or is that a known issue fixed in a later release? Thanks for any insight. David Morel -- Yiftach Shoolman +972-54-7634621 -- Yiftach Shoolman +972-54-7634621 David Morel -- Booking.com http://www.booking.com/ Lyon office: 93 rue de la Villette, 5e étage, F-69003 Lyon phone:+33 4 20 10 26 63 gsm:+33 6 80 38 56 83 -- Yiftach Shoolman +972-54-7634621
Re: Some set() calls taking way too long
Have you seen/gone through this: http://memcached.org/timeouts checked the common stuff on the page (not swapping a little bit or under memory pressure/etc), and used the independent connection tester tool to validate that just sets instead of gets or new connections are being problematic? Going through the motions there helps rule out a lot of stuff. It'd be great if you were on a newer version as well, but that may not solve any problems here. With four threads, if a set blocks for 10 milliseconds due to a memcached related issue, 1/4th of all your gets would also block for at least 10 milliseconds. Since you're not seeing that I'm willing to bet it's something else? unless that's not true, in which the testing above should confirm it. Also need to know what client you're using, in case you're on one which batches sets, or the bug is in the client. On Tue, 17 Jul 2012, David Morel wrote: On 17 Jul 2012, at 15:33, Yiftach Shoolman wrote: Re #2 - when more objects are stored in the cache, the hit ratio should be higher-- the application might run faster, i.e. with less DB accesses. no relation here, I was really mentioning bare set() calls timing, regardless of anything else. Anyway, it sounds to me more related to slabs allocation, as only reset solves it (not flush) if I understood u well. all the slabs are already allocated, with a moderate rate of evictions. so it's not slab allocation either, it's just one key now and then, regardless (it seems) of the key itself or the size of the object. Does it happen on any object size or on specific object size range ? anything, really. puzzling, hey? thanks! On Tue, Jul 17, 2012 at 4:04 PM, David Morel david.mo...@amakuru.netwrote: On Tuesday, July 17, 2012 12:26:16 PM UTC+2, Yiftach wrote: Few things that may help understanding your problem: 1. What is the status of your slabs allocation, is there enough room to all slabes ? This happens when the memory gets close to full. however there is not a large number of evictions. I would expect evictions to be made whenever needed, but not the process of making room for 1 object to take half a second. 2. Do you see increase in the requests rate when your Memcached memory is becoming full with objects ? I don't think so, why would that be the case, it's application dependent, not server, right? 3. How many threads are configured ? the default 4 On Tue, Jul 17, 2012 at 1:11 PM, David Morel david.mo...@amakuru.netwrote: hi memcached users/devvers, I'm seeing occasional slowdowns (tens of milliseconds) in setting some keys on some big servers (80GB RAM allocated to memcached) which contain a large number of keys (many millions). The current version I use is 1.4.6 on RH6. The thing is once I bounce the service (restart, not flush_all), everything becomes fine again. So could a large number of keys be the source of the issue (some memory allocation slowdown or something)? I don't see that many evictions on the box, and anyway, evicting an object to make room for another shouldn't take long, should it? Is there a remote possibility the large number of keys is at fault and splitting the daemons, like 2 or more instances per box, would fix it? Or is that a known issue fixed in a later release? Thanks for any insight. David Morel -- Yiftach Shoolman +972-54-7634621 -- Yiftach Shoolman +972-54-7634621 David Morel -- Booking.com http://www.booking.com/ Lyon office: 93 rue de la Villette, 5e étage, F-69003 Lyon phone:+33 4 20 10 26 63 gsm:+33 6 80 38 56 83