Hi Garo, I haven't had this issue on SSDs, but I have definitely seen it with spinning drives. I would think that SSDs would have more than enough bandwidth to keep up with requests, but you may be running into issues with Cassandra calling fsync on the commitlog.
What are your settings for the following? commitlog_sync commitlog_sync_period_in_ms commitlog_sync_batch_window_in_ms If you're using periodic, you could try changing commitlog_sync_period_in_ms to something smaller like 1000 ms and seeing if the problem is reduced (the theory is that there would be less pending data to sync). If you are using batch, switch to periodic. You could try mounting a GP2 volume and putting the commit log directory on it and see if the problem goes away (say 200 GB for sufficient IOPS). I'm guessing you don't have much in the way of unallocated blocks in your LVM vg. Writing to the commit log is single threaded, and if the commit log is tied up waiting for IO during an fsync, it will block writes to the node. If the threads are blocked on writing, the nodes will also be stall for reading. The symptoms you are seeing are exactly the same as I saw with spinning rust. I'm not sure why you didn't see this problem with EBS. -Mark On Sat, Jul 23, 2016 at 7:21 AM, Juho Mäkinen <juho.maki...@gmail.com> wrote: > Hi Mark. > > I have an LVM volume which stripes the four ephemeral SSD:s in the system > and we use that for both data and commit log. I've used similar setup in the > past (but with EBS) and we didn't see this behavior. Each node gets just > around 250 writes per second. It is possible that the commit log is the > issue here, but could I somehow measure it from the JMX metrics without the > need of restructuring my entire cluster? > > Here's a screenshot from the latencies from our application point of view, > which uses the Cassandra cluster to do reads. I started a rolling restart at > around 09:30 and you can clearly see how the system latency dropped. > http://imgur.com/a/kaPG7 > > On Sat, Jul 23, 2016 at 2:25 AM, Mark Rose <markr...@markrose.ca> wrote: >> >> Hi Garo, >> >> Did you put the commit log on its own drive? Spiking CPU during stalls >> is a symptom of not doing that. The commitlog is very latency >> sensitive, even under low load. Do be sure you're using the deadline >> or noop scheduler for that reason, too. >> >> -Mark >> >> On Fri, Jul 22, 2016 at 4:44 PM, Juho Mäkinen <juho.maki...@gmail.com> >> wrote: >> >> Are you using XFS or Ext4 for data? >> > >> > >> > We are using XFS. Many nodes have a couple large SSTables (in order of >> > 20-50 >> > GiB), but I havent cross checked if the load spikes happen only on >> > machines >> > which have these tables. >> > >> >> >> >> As an aside, for the amount of reads/writes you're doing, I've found >> >> using c3/m3 instances with the commit log on the ephemeral storage and >> >> data on st1 EBS volumes to be much more cost effective. It's something >> >> to look into if you haven't already. >> > >> > >> > Thanks for the idea! I previously used c4.4xlarge instances with two >> > 1500 GB >> > GP2 volumes, but I found out that we maxed out their bandwidth too >> > easily, >> > so that's why my newest cluster is based on i2.4xlarge instances. >> > >> > And to answer Ryan: No, we are not using counters. >> > >> > I was thinking that could the big amount (100+ GiB) of mmap'ed files >> > somehow >> > cause some inefficiencies on the kernel side. That's why I started to >> > learn >> > on kernel huge pages and came up with the idea of disabling the huge >> > page >> > defrag, but nothing what I've found indicates that this can be a real >> > problem. After all, Linux fs cache is a really old feature, so I expect >> > it >> > to be pretty bug free. >> > >> > I guess that I have to next learn how the load value itself is >> > calculated. I >> > know about the basic idea that when load is below the number of CPUs >> > then >> > the system should still be fine, but there's at least the iowait which >> > is >> > also used to calculate the load. So because I am not seeing any >> > extensive >> > iowait, and my userland CPU usage is well below what my 16 cores should >> > handle, then what else contributes to the system load? Can I somehow >> > make >> > any educated guess what the high load might tell me if it's not iowait >> > and >> > it's not purely userland process CPU usage? This is starting to get >> > really >> > deep really fast :/ >> > >> > - Garo >> > >> > >> >> >> >> >> >> -Mark >> >> >> >> On Fri, Jul 22, 2016 at 8:10 AM, Juho Mäkinen <juho.maki...@gmail.com> >> >> wrote: >> >> > After a few days I've also tried disabling Linux kernel huge pages >> >> > defragement (echo never > /sys/kernel/mm/transparent_hugepage/defrag) >> >> > and >> >> > turning coalescing off (otc_coalescing_strategy: DISABLED), but >> >> > either >> >> > did >> >> > do any good. I'm using LCS, there are no big GC pauses, and I have >> >> > set >> >> > "concurrent_compactors: 5" (machines have 16 CPUs), but there are >> >> > usually >> >> > not any compactions running when the load spike comes. "nodetool >> >> > tpstats" >> >> > shows no running thread pools except on the Native-Transport-Requests >> >> > (usually 0-4) and perhaps ReadStage (usually 0-1). >> >> > >> >> > The symptoms are the same: after about 12-24 hours increasingly >> >> > number >> >> > of >> >> > nodes start to show short CPU load spikes and this affects the median >> >> > read >> >> > latencies. I ran a dstat when a load spike was already under way (see >> >> > screenshot http://i.imgur.com/B0S5Zki.png), but any other column than >> >> > the >> >> > load itself doesn't show any major change except the system/kernel >> >> > CPU >> >> > usage. >> >> > >> >> > All further ideas how to debug this are greatly appreciated. >> >> > >> >> > >> >> > On Wed, Jul 20, 2016 at 7:13 PM, Juho Mäkinen >> >> > <juho.maki...@gmail.com> >> >> > wrote: >> >> >> >> >> >> I just recently upgraded our cluster to 2.2.7 and after turning the >> >> >> cluster under production load the instances started to show high >> >> >> load >> >> >> (as >> >> >> shown by uptime) without any apparent reason and I'm not quite sure >> >> >> what >> >> >> could be causing it. >> >> >> >> >> >> We are running on i2.4xlarge, so we have 16 cores, 120GB of ram, >> >> >> four >> >> >> 800GB SSDs (set as lvm stripe into one big lvol). Running >> >> >> 3.13.0-87-generic >> >> >> on HVM virtualisation. Cluster has 26 TiB of data stored in two >> >> >> tables. >> >> >> >> >> >> Symptoms: >> >> >> - High load, sometimes up to 30 for a short duration of few >> >> >> minutes, >> >> >> then >> >> >> the load drops back to the cluster average: 3-4 >> >> >> - Instances might have one compaction running, but might not have >> >> >> any >> >> >> compactions. >> >> >> - Each node is serving around 250-300 reads per second and around >> >> >> 200 >> >> >> writes per second. >> >> >> - Restarting node fixes the problem for around 18-24 hours. >> >> >> - No or very little IO-wait. >> >> >> - top shows that around 3-10 threads are running on high cpu, but >> >> >> that >> >> >> alone should not cause a load of 20-30. >> >> >> - Doesn't seem to be GC load: A system starts to show symptoms so >> >> >> that >> >> >> it >> >> >> has ran only one CMS sweep. Not like it would do constant >> >> >> stop-the-world >> >> >> gc's. >> >> >> - top shows that the C* processes use 100G of RSS memory. I assume >> >> >> that >> >> >> this is because cassandra opens all SSTables with mmap() so that >> >> >> they >> >> >> will >> >> >> pop up in the RSS count because of this. >> >> >> >> >> >> What I've done so far: >> >> >> - Rolling restart. Helped for about one day. >> >> >> - Tried doing manual GC to the cluster. >> >> >> - Increased heap from 8 GiB with CMS to 16 GiB with G1GC. >> >> >> - sjk-plus shows bunch of SharedPool workers. Not sure what to make >> >> >> of >> >> >> this. >> >> >> - Browsed over >> >> >> https://tobert.github.io/pages/als-cassandra-21-tuning-guide.html >> >> >> but >> >> >> didn't >> >> >> find any apparent >> >> >> >> >> >> I know that the general symptom of "system shows high load" is not >> >> >> very >> >> >> good and informative, but I don't know how to better describe what's >> >> >> going >> >> >> on. I appreciate all ideas what to try and how to debug this >> >> >> further. >> >> >> >> >> >> - Garo >> >> >> >> >> > >> > >> > > >
