Apart from all that you can try to reduce the compression chunk size from the default 64kb to 16kb or even down to 4kb. This can help a lot if your read io on disk is very high and the page cache is not efficient.
Am 21.07.2017 23:03 schrieb "Petrus Gomes" <petru...@gmail.com>: > Thanks a lot to share the result. > > Boa Sorte. > ;-) > Take care. > Petris Silva > > On Fri, Jul 21, 2017 at 12:19 PM, Felipe Esteves < > felipe.este...@b2wdigital.com> wrote: > >> Hi, Petrus, >> >> Seems we've solved the problem, but it wasn't relationed to repair the >> cluster or disk latency. >> I've increased the memory available for Cassandra from 16GB to 24GB and >> the performance was much improved! >> The main symptom we've observed in Opscenter was a significantly decrease >> in total compactions graph. >> >> Felipe Esteves >> >> Tecnologia >> >> felipe.este...@b2wdigital.com <seu.em...@b2wdigital.com> >> >> >> >> 2017-07-15 3:23 GMT-03:00 Petrus Gomes <petru...@gmail.com>: >> >>> Hi Felipe, >>> >>> Yes, try it and let us know how it goes. >>> >>> Thanks, >>> Petrus Silva. >>> >>> On Fri, Jul 14, 2017 at 11:37 AM, Felipe Esteves < >>> felipe.este...@b2wdigital.com> wrote: >>> >>>> Hi Petrus, thanks for the feedback. >>>> >>>> I couldn't found the percent repaired in nodetool info, C* version is >>>> 2.1.8, maybe it's something newer than that? >>>> >>>> I'm analyzing this thread about num_token. >>>> >>>> Compaction is "compaction_throughput_mb_per_sec: 16", I don't get >>>> pending compactions in Opscenter. >>>> >>>> One point I've noticed, is that Opscenter show "OS: Disk Latency" max >>>> with high values when the problem occurs, but it doesn't reflect in server >>>> directly monitoring, in these tools the IO and latency of disks seems ok. >>>> But seems to me that "read repair attempted" is a bit high, maybe it >>>> will explain the latency in reads. I will try to run a repair on cluster to >>>> see how it goes. >>>> >>>> Felipe Esteves >>>> >>>> Tecnologia >>>> >>>> felipe.este...@b2wdigital.com <seu.em...@b2wdigital.com> >>>> >>>> Tel.: (21) 3504-7162 ramal 57162 >>>> >>>> Skype: felipe2esteves >>>> >>>> 2017-07-13 15:02 GMT-03:00 Petrus Gomes <petru...@gmail.com>: >>>> >>>>> How is your Percent Repaired when you run " nodetool info" ? >>>>> >>>>> Search for : >>>>> "reduced num_token = improved performance ??" topic. >>>>> The people were discussing that. >>>>> >>>>> How is your compaction is configured? >>>>> >>>>> Could you run the same process in command line to have a measurement? >>>>> >>>>> Thanks, >>>>> Petrus Silva >>>>> >>>>> >>>>> >>>>> On Thu, Jul 13, 2017 at 7:49 AM, Felipe Esteves < >>>>> felipe.este...@b2wdigital.com> wrote: >>>>> >>>>>> Hi, >>>>>> >>>>>> I have a Cassandra 2.1 cluster running on AWS that receives high read >>>>>> loads, jumping from 100k requests to 400k requests, for example. Then it >>>>>> normalizes and later cames another high throughput. >>>>>> >>>>>> To the application, it appears that Cassandra is slow. However, cpu >>>>>> and disk use is ok in every instance, row cache is enabled and with >>>>>> almost >>>>>> 100% hit rate. >>>>>> >>>>>> The logs from Cassandra instances doesn't have any errors, nor >>>>>> tombstone messages or something liked that. It's mostly compactions and >>>>>> G1GC operations. >>>>>> >>>>>> Any hints on where to investigate more? >>>>>> >>>>>> >>>>>> Felipe Esteves >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>> >>>>> ------------------------------ >>>>> >>>>> Esta mensagem pode conter informações confidenciais e somente o >>>>> indivíduo ou entidade a quem foi destinada pode utilizá-la. A transmissão >>>>> incorreta da mensagem não acarreta a perda de sua confidencialidade. Caso >>>>> esta mensagem tenha sido recebida por engano, solicitamos que o fato seja >>>>> comunicado ao remetente e que a mensagem seja eliminada de seu sistema >>>>> imediatamente. É vedado a qualquer pessoa que não seja o destinatário >>>>> usar, >>>>> revelar, distribuir ou copiar qualquer parte desta mensagem. Ambiente de >>>>> comunicação sujeito a monitoramento. >>>>> >>>>> This message may include confidential information and only the >>>>> intended addresses have the right to use it as is, or any part of it. A >>>>> wrong transmission does not break its confidentiality. If you've received >>>>> it because of a mistake or erroneous transmission, please notify the >>>>> sender >>>>> and delete it from your system immediately. This communication environment >>>>> is controlled and monitored. >>>>> >>>>> B2W Digital >>>>> >>>>> >>>>> >>>> >>>> >>>> >>>> >>> >> >> >> >