# Low writing speed on SSD, identify bottleneck **URL:** \ **Category:** Elasticsearch\ **Created:** [October 2, 2021, 5:25am UTC](https://discuss.elastic.co/t/low-writing-speed-on-ssd-identify-bottleneck/285726 "2021-10-02T05:25:34Z")\ **Posts on this page:** 7\ **Page:** 1
**Author:** ![sdwerwed](https://sea2.discourse-cdn.com/elastic/user_avatar/discuss.elastic.co/sdwerwed/32/95363_2.png) [@sdwerwed](https://discuss.elastic.co/u/sdwerwed)\ **Post date:** [October 2, 2021, 5:25am UTC](https://discuss.elastic.co/t/low-writing-speed-on-ssd-identify-bottleneck/285726/1 "2021-10-02T05:25:34Z")
Hello Elasticseach community! 🙂 Writing speed on SSD is low and I am not sure where the bottleneck is. - I use Elasticsearch 7.14.0 in Kubernetes - 4 data nodes - 3 masters - 8G RAM 4g heap 1 CPU each - 512 Standart SSD each. I did esrally event data tests with various bulk sizes (see blue annotations), the maximum write speed per node was 20MB/s, and max 30K docs/s in total, CPU is very low and memory is used normally. I have the feeling that there is a big bottleneck somewhere because disk in not saturated (or do you think the opposite I am not sure about that) Index rate (in stack) 4 nodes contribute to 30k logs/s: ![image](https://us1.discourse-cdn.com/elastic/original/3X/7/a/7a7b22be5bd7ba3ad0943123260a418786b46cf1.png) Disk metrics are better with higher bulk size: ![image](https://us1.discourse-cdn.com/elastic/original/3X/5/c/5c2a917661f364ea5a3b98e11923645bf9bdb375.png) I think low block size + fsync might affect the writing speed on SSD disks a lot, it may be the reason for slow writing(?). I tried that in many different VMs and the results are kind of same. I am able to simulate something similar with fio. With block\_size=4k and fsync=1 it gives 12MB/s, if block\_size=4000 and fsync=40 SSD is fast as expected. Any idea how to set the block\_size or similar setting to saturate the disk so it can handle more logs/s? Here are the results of fio: ```auto fio --name TEST --eta-newline=5s --filename=fio-tempfile.dat --rw=write --size=500m --io_size=10g --blocksize=200k --ioengine=libaio --fsync=1 --iodepth=1 --direct=1 --numjobs=1 --runtime=60 --group_reporting TEST: (g=0): rw=write, bs=(R) 200KiB-200KiB, (W) 200KiB-200KiB, (T) 200KiB-200KiB, ioengine=libaio, iodepth=1 fio-3.1 Starting 1 process Jobs: 1 (f=1): [W(1)][11.7%][r=0KiB/s,w=17.6MiB/s][r=0,w=90 IOPS][eta 00m:53s] Jobs: 1 (f=1): [W(1)][21.7%][r=0KiB/s,w=4400KiB/s][r=0,w=22 IOPS][eta 00m:47s] Jobs: 1 (f=1): [W(1)][31.7%][r=0KiB/s,w=13.1MiB/s][r=0,w=67 IOPS][eta 00m:41s] Jobs: 1 (f=1): [W(1)][41.7%][r=0KiB/s,w=9.78MiB/s][r=0,w=50 IOPS][eta 00m:35s] Jobs: 1 (f=1): [W(1)][51.7%][r=0KiB/s,w=13.5MiB/s][r=0,w=69 IOPS][eta 00m:29s] Jobs: 1 (f=1): [W(1)][62.3%][r=0KiB/s,w=8000KiB/s][r=0,w=40 IOPS][eta 00m:23s] Jobs: 1 (f=1): [W(1)][72.1%][r=0KiB/s,w=12.1MiB/s][r=0,w=62 IOPS][eta 00m:17s] Jobs: 1 (f=1): [W(1)][82.0%][r=0KiB/s,w=16.0MiB/s][r=0,w=87 IOPS][eta 00m:11s] Jobs: 1 (f=1): [W(1)][91.8%][r=0KiB/s,w=5400KiB/s][r=0,w=27 IOPS][eta 00m:05s] Jobs: 1 (f=1): [W(1)][100.0%][r=0KiB/s,w=15.4MiB/s][r=0,w=79 IOPS][eta 00m:00s] TEST: (groupid=0, jobs=1): err= 0: pid=4978: Sat Oct 2 01:30:49 2021 write: IOPS=57, BW=11.3MiB/s (11.8MB/s)(677MiB/60004msec) slat (usec): min=151, max=9712, avg=305.04, stdev=399.82 clat (usec): min=2, max=259, avg= 3.44, stdev= 5.35 lat (usec): min=154, max=9719, avg=309.68, stdev=400.23 clat percentiles (usec): | 1.00th=[3], 5.00th=[3], 10.00th=[3], 20.00th=[3], | 30.00th=[3], 40.00th=[3], 50.00th=[4], 60.00th=[4], | 70.00th=[4], 80.00th=[4], 90.00th=[4], 95.00th=[5], | 99.00th=[6], 99.50th=[21], 99.90th=[71], 99.95th=[86], | 99.99th=[260] bw ( KiB/s): min= 3585, max=18800, per=99.98%, avg=11556.20, stdev=5193.10, samples=120 iops : min= 17, max= 94, avg=57.66, stdev=25.96, samples=120 lat (usec) : 4=94.87%, 10=4.35%, 20=0.23%, 50=0.40%, 100=0.12% lat (usec) : 500=0.03% cpu : usr=0.16%, sys=0.62%, ctx=7379, majf=0, minf=11 IO depths : 1=100.0%, 2=0.0%, 4=0.0%, 8=0.0%, 16=0.0%, 32=0.0%, >=64=0.0% submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0% complete : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0% issued rwt: total=0,3468,0, short=0,0,0, dropped=0,0,0 latency : target=0, window=0, percentile=100.00%, depth=1 Run status group 0 (all jobs): WRITE: bw=11.3MiB/s (11.8MB/s), 11.3MiB/s-11.3MiB/s (11.8MB/s-11.8MB/s), io=677MiB (710MB), run=60004-60004msec ``` Update: @jprante had commented something related but it was 8 years ago, maybe something have changed from then... > [@Recommended File System Settings for Elasticsearch](https://discuss.elastic.co/t/recommended-file-system-settings-for-elasticsearch/18733/4): > > Elasticsearch does not set the read/write in 4k blocks, this is up to the JVM/OS. Elastisearch uses Lucene and Lucene wraps all reads and write into Java I/O streams, where Java uses the OS layer, and the OS uses the file system drivers. There are some rule of thumb: most important, don't mess with the block size. Because the virtual memory page size of Linux is 4k, and reading/writing pages from devices to memory should be fast, you will often find 4k blocksize and a factor of 4k for buf… ---
**Author:** ![Christian\_Dahlqvist](https://sea2.discourse-cdn.com/elastic/user_avatar/discuss.elastic.co/christian_dahlqvist/32/4617_2.png) [@Christian\_Dahlqvist](https://discuss.elastic.co/u/Christian_Dahlqvist)\ **Post date:** [October 2, 2021, 6:06am UTC](https://discuss.elastic.co/t/low-writing-speed-on-ssd-identify-bottleneck/285726/2 "2021-10-02T06:06:49Z")
1 CPU sounds very low. Can you show CPU utilisation? How many indices and shards are you actively indexing into? Are you indexing immutable documents and letting Elasticsearch assign the IDs or are you doing something else? Where is this hosted? What type of SSD are you using? Is it locally attached to the hosts? ---
**Author:** ![sdwerwed](https://sea2.discourse-cdn.com/elastic/user_avatar/discuss.elastic.co/sdwerwed/32/95363_2.png) [@sdwerwed](https://discuss.elastic.co/u/sdwerwed)\ **Post date:** [October 2, 2021, 12:48pm UTC](https://discuss.elastic.co/t/low-writing-speed-on-ssd-identify-bottleneck/285726/3 "2021-10-02T12:48:04Z")
First of all, thank you Christian for your reply! - I do not have the metrics right now but I have set also many vCPUs and did not change the result, utilization was below 40% but might try again today. - Shards were around 5. - esrally creates the index that by: `esrally race --track=eventdata --target-hosts=elasticsearch-data:9200 --pipeline=benchmark-only --client-options="timeout:60,use_ssl:false" --include-tasks="delete-index,create-index,index-append" --kill-running-processes --track-params="bulk_size:10000,ingest_percentage:100,bulk_indexing_clients:8" ` I think is this [index.json](https://github.com/elastic/rally-tracks/blob/master/eventdata/index.json) - It is attached but I have tried the SSD OS use and was the same, the SSD type is [Standard SSD E20](https://docs.microsoft.com/en-us/azure/virtual-machines/disks-types#standard-ssd), in my laptop the fio with the specific settings do not saturate the SSD too Any idea? 🙂 ---
**Author:** ![Christian\_Dahlqvist](https://sea2.discourse-cdn.com/elastic/user_avatar/discuss.elastic.co/christian_dahlqvist/32/4617_2.png) [@Christian\_Dahlqvist](https://discuss.elastic.co/u/Christian_Dahlqvist)\ **Post date:** [October 2, 2021, 2:28pm UTC](https://discuss.elastic.co/t/low-writing-speed-on-ssd-identify-bottleneck/285726/4 "2021-10-02T14:28:23Z")
Are you running this whole cluster on a laptop? If so, what is the specification? Elasticsearch does fsync frequently during indexing so indexing to a lot if shards with replicas can result in a lot if small batches being synced. Having larger batches would increase the number of documents per batch and therefore reduce the number of fsyncs per data volume. ---
**Author:** ![sdwerwed](https://sea2.discourse-cdn.com/elastic/user_avatar/discuss.elastic.co/sdwerwed/32/95363_2.png) [@sdwerwed](https://discuss.elastic.co/u/sdwerwed)\ **Post date:** [October 2, 2021, 2:38pm UTC](https://discuss.elastic.co/t/low-writing-speed-on-ssd-identify-bottleneck/285726/5 "2021-10-02T14:38:37Z")
Hello, looks like the higher vCPUs helped, I had tried high CPU but the low bulk size and maybe saturated the disk that's why CPU was underutilized, but the high bulk size and high CPU runs better, I reached 130K logs/s. I run in AKS, but tested fio in my laptop to see if there is a problem with Azure disks but there is no. ![image](https://us1.discourse-cdn.com/elastic/original/3X/c/a/ca430d9fd8af4bd0d3cebc2542258c1a475f6d19.png) ---
**Author:** ![sdwerwed](https://sea2.discourse-cdn.com/elastic/user_avatar/discuss.elastic.co/sdwerwed/32/95363_2.png) [@sdwerwed](https://discuss.elastic.co/u/sdwerwed)\ **Post date:** [October 2, 2021, 3:35pm UTC](https://discuss.elastic.co/t/low-writing-speed-on-ssd-identify-bottleneck/285726/6 "2021-10-02T15:35:19Z")
With 8 vCPU per data node, the write speed is **50MB/s and 200IOPS** per node, isn't that still low? Setting index.refresh\_interval to 1s, -1, 30s did not change the write bandwidth. The total open file descriptors per node are 1.44 K, 7 primary shards, 8 total shards, around 151 segments with 1KB-3G sizes, 3 data nodes, 8 vCPU 16RAM, 8g heap. **Writing size is around 300-400KB** , any idea how to maximize the write speed by changing the block size (which will result in less IOPS) without increasing too much the CPU? ![image](https://us1.discourse-cdn.com/elastic/original/3X/8/3/8384c9cb5bc2bbf541fb93d98a33c8c1e6203cd2.png) Durring esrally I get those warnings ```auto [WARNING] merges_total_time is 5829883 ms indicating that the cluster is not in a defined clean state. Recorded index time metrics may be misleading. [WARNING] merges_total_throttled_time is 3772726 ms indicating that the cluster is not in a defined clean state. Recorded index time metrics may be misleading. [WARNING] indexing_total_time is 8133911 ms indicating that the cluster is not in a defined clean state. Recorded index time metrics may be misleading. [WARNING] refresh_total_time is 332101 ms indicating that the cluster is not in a defined clean state. Recorded index time metrics may be misleading. [WARNING] flush_total_time is 766505 ms indicating that the cluster is not in a defined clean state. Recorded index time metrics may be misleading. ``` ---
**Author:** ![system](https://us1.discourse-cdn.com/elastic/original/3X/1/a/1ac57faf039f6b580b3f104ef42a2a89e41014de.png) [@system](https://discuss.elastic.co/u/system)\ **Post date:** [October 30, 2021, 3:35pm UTC](https://discuss.elastic.co/t/low-writing-speed-on-ssd-identify-bottleneck/285726/7 "2021-10-30T15:35:44Z")
This topic was automatically closed 28 days after the last reply. New replies are no longer allowed.