We have 2 node clusters each node having 32 core + 198 GB of memory + locally attached storage.
Running a search query ( for around 15 minutes ) all CPU hits 100% CPU.
I can infer from iostat that I/O is not bottleneck ( %io-utilization is 0 and io queue length is 0 ) and Heap size never goes beyond 2 GB. JVM heap allocated is around 24 GB. Number of search threads 1re 96 ( Num-of-Cores * 3 )
Any suggestion/recommendation, where the bottleneck is or is this behavior is expected ?
We are using ES 1.2.x
Thanks
Raj
Related
We have been running ProxmoxVE since 5.0 (now in 6.4-15) and we noticed a decay in performance whenever there is some heavy reading/writing.
We have 9 nodes, 7 with CEPH and 56 OSDs (8 on each node). OSDs are hard drives (HDD) WD Gold or better (4~12 Tb). Nodes with 64/128 Gbytes RAM, dual Xeon CPU mainboards (various models).
We already tried simple tests like "ceph tell osd.* bench" getting stable 110 Mb/sec data transfer to each of them with +- 10 Mb/sec spread during normal operations. Apply/Commit Latency is normally below 55 ms with a couple of OSDs reaching 100 ms and one-third below 20 ms.
The front network and back network are both 1 Gbps (separated in VLANs), we are trying to move to 10 Gbps but we found some trouble we are still trying to figure out how to solve (unstable OSDs disconnections).
The Pool is defined as "replicated" with 3 copies (2 needed to keep running). Now the total amount of disk space is 305 Tb (72% used), reweight is in use as some OSDs were getting much more data than others.
Virtual machines run on the same 9 nodes, most are not CPU intensive:
Avg. VM CPU Usage < 6%
Avg. Node CPU Usage < 4.5%
Peak VM CPU Usage 40%
Peak Node CPU Usage 30%
But I/O Wait is a different story:
Avg. Node IO Delay 11
Max. Node IO delay 38
Disk writing load is around 4 Mbytes/sec average, with peaks up to 20 Mbytes/sec.
Anyone with experience in getting better Proxmox+CEPH performance?
Thank you all in advance for taking the time to read,
Ruben.
Got some Ceph pointers that you could follow...
get some good NVMEs (one or two per server but if you have 8HDDs per server 1 should be enough) and put those as DB/WALL (make sure they have power protection)
the ceph tell osd.* bench is not that relevant for real world, I suggest to try some FIO tests see here
set OSD osd_memory_target to at 8G or RAM minimum.
in order to save some write on your HDD (data is not replicated X times) create your RBD pool as EC (erasure coded pool) but please do some research on that because there are some tradeoffs. Recovery takes some extra CPU calculations
All and all, hype-converged clusters are good for training, small projects and medium projects with not such a big workload on them... Keep in mind that planning is gold
Just my 2 cents,
B.
Looking at Jeff Dean's famous latency guides
Latency Comparison Numbers (~2012)
----------------------------------
L1 cache reference 0.5 ns
Branch mispredict 5 ns
L2 cache reference 7 ns 14x L1 cache
Mutex lock/unlock 25 ns
Main memory reference 100 ns 20x L2 cache, 200x L1 cache
Compress 1K bytes with Zippy 3,000 ns 3 us
Send 1K bytes over 1 Gbps network 10,000 ns 10 us
Read 4K randomly from SSD* 150,000 ns 150 us ~1GB/sec SSD
Read 1 MB sequentially from memory 250,000 ns 250 us
Round trip within same datacenter 500,000 ns 500 us
Read 1 MB sequentially from SSD* 1,000,000 ns 1,000 us 1 ms ~1GB/sec SSD, 4X memory
Disk seek 10,000,000 ns 10,000 us 10 ms 20x datacenter roundtrip
Read 1 MB sequentially from disk 20,000,000 ns 20,000 us 20 ms 80x memory, 20X SSD
Send packet CA->Netherlands->CA 150,000,000 ns 150,000 us 150 ms
One thing which looks somewhat uncanny to me is the time taken to read 1MB sequentially from disk is only 10 times faster than sending a round trip packet across the Atlantic. Can anyone give me more intuition why this feels right.
Q : 1MB SEQ-HDD-READ ~ 10x faster than a CA/NL trans-atlantic RTT - why this feels right?
Some "old" values ( with a few cross-QPI/NUMA updates from 2017 ) to start from:
0.5 ns - CPU L1 dCACHE reference
1 ns - speed-of-light (a photon) travel a 1 ft (30.5cm) distance
5 ns - CPU L1 iCACHE Branch mispredict
7 ns - CPU L2 CACHE reference
71 ns - CPU cross-QPI/NUMA best case on XEON E5-46*
100 ns - MUTEX lock/unlock
100 ns - CPU own DDR MEMORY reference
135 ns - CPU cross-QPI/NUMA best case on XEON E7-*
202 ns - CPU cross-QPI/NUMA worst case on XEON E7-*
325 ns - CPU cross-QPI/NUMA worst case on XEON E5-46*
10,000 ns - Compress 1 KB with Zippy PROCESS (+GHz,+SIMD,+multicore tricks)
20,000 ns - Send 2 KB over 1 Gbps NETWORK
250,000 ns - Read 1 MB sequentially from MEMORY
500,000 ns - Round trip within a same DataCenter
10,000,000 ns - DISK seek
10,000,000 ns - Read 1 MB sequentially from NETWORK
30,000,000 ns - Read 1 MB sequentially from DISK
150,000,000 ns - Send a NETWORK packet CA -> Netherlands
| | | |
| | | ns|
| | us|
| ms|
Trans-Atlantic Network RTT :
Global optical networks work roughly at a speed of light ( 300.000.000 m/s )
LA(CA)-AMS(NL) packet has to travel not the geodetical "distance", but over a set of continental and trans-atlantic "submarine" cables, the length of which is way longer ( see the map )
These factors do not "improve" - only the transport capacity is growing, with add-on latencies introduced in light-amplifiers, retiming units and other L1-PHY / L2-/L3-networking technologies are kept under control, as small as possible.
So the LA(CA)-AMS(NL) RTT will remain, using this technology, the same ~ 150 ms
Using other technology, LEO-Sat Cubes - as an example - the "distance" will only grow from ~ 9000 km P2P, by a pair of additional GND/LEO segments, plus by a few addition LEO/LEO hops, which introduce "longer" distance, add-on hop/hop re-processing latencies and capacity will not get any close to the current optical transports available, so no magic jump "back to the future" is to be expected ( we still miss the DeLorean ).
The HDD Disk :
HDD-s can have very fast and very short transport-path for moving the data, but the READ-ops have to wait for the physical / mechanical operations of the media-reading heads ( that takes most of the time here, not the actual data-transfer to the host RAM )
HDD-s are rotational devices, the disk has to "align" where to start the read, which costs the first about 10 [ms]
HDD-s devices store data into a static structure of heads( 2+, reading physical signals from the magnetic plates' surfaces ):cylinders( concentric circular zones on the plate, into which a cyl-aligned reading-head gets settled by disk-head micro-controller):sector( angular-sections of the cylinder, each carrying a block of the same sized data ~ 4KB, 8KB, ... )
These factors do not "improve" - all commodity produced drives remain at industry selected angular speeds of about { 5k4 | 7k2 | 10k | 15k | 18k }-spins/min (RPM). This means, that if a well-compacted data-layouts are maintained on such a disk, one continuous head:cylinder aligned reading round the whole cylinder will take:
>>> [ 1E3 / ( RPM / 60. ) for RPM in ( 5400, 7200, 10000, 15000, 18000 ) ]
11.1 ms per CYL # 5k4 RPM disk,
8.3 ms per CYL # 7k2 RPM disk,
6.0 ms per CYL # 10k RPM disk,
4.0 ms per CYL # 15k RPM disk,
3.3 ms per CYL # 18k RPM disk.
Data-density is also limited by the magnetic media properties. Spintronics R&D will bring some more densely stored data, yet the last 30 years have been well inside the limits of the reliable magnetic storage.
More is to be expected from a trick to co-parallel-read from several heads at-once, yet this goes against the design of the embedded microcontrollers, so most of the reading goes but sequentially, from one head after another, into the HDD-controller onboard buffers, best if no cyl-to-cyl heads mechanical re-alignment were to take place ( technically this depends on the prior data-to-disc layout, maintained by the O/S and possible care of disk-optimisers ( originally called disk disk-"compression", which just tried to re-align the known sequences of FAT-described data-blocks, so as to follow the most optimal trajectory of head:cyl:sector transitions, depending most on the actual device's head:head and cyl:cyl latencies ). So even the most optimistic data-layout takes ~ 13..21 [ms] to seek-and-read but one head:cyl-path
Laws of Physics decide
Some numbers from 2020.
Load from L1 is 4 cycles on Intel Coffee Lake and Ryzen (0.8nsec on a 5GHz CPU).
Load from memory is ~215 cycles on Intel Coffee Lake (43nsec on a 5GHz CPU). ~280 cycles on Ryzen.
I have a cluster of 13 machines with 4 physical CPUs and 24 G of RAM.
I started a spark cluster with one driver and 12 slaves.
I set the number of cores by slaves to 12 cores, meaning I have a cluster as foloowing :
Alive Workers: 12
Cores in use: 144 Total, 110 Used
Memory in use: 263.9 GB Total, 187.0 GB Used
I started an application with the folowing configuration :
[('spark.driver.cores', '4'),
('spark.executor.memory', '15G'),
('spark.executor.id', 'driver'),
('spark.driver.memory', '5G'),
('spark.python.worker.memory', '1042M'),
('spark.cores.max', '96'),
('spark.rdd.compress', 'True'),
('spark.serializer.objectStreamReset', '100'),
('spark.executor.cores', '8'),
('spark.default.parallelism', '48')]
I understand there are 15G of RAM by executor with 8 task slot and a parallelism of 48 (48 = 6 task slot * 12 slaves).
then I have two big files on HDFS : 6 G each, (from a directory of 12 files of 5 blocks of 128 Mb each) , with a 3x replication factor.
I union these two files => I get one dataframe of 12 GB I think but I see a 37 G reading input through the IHM :
That could be the first question : Why 37 Gb ?
Then as the execution time is too long for me, I try to cache the data so that I can go faster. But the caching method never finishes, here you can see it is already 45 minutes before the end (Vs 6 min not cached !):
So I try to understand why, and I see the usage of Memory/Disk on the storage section of the ihm :
So there are some part of the RDD that are staying on disk.
Furthemore I see the executors may still have free memory :
And I notice on the same "storage" page that the size of the RDD has jumped :
Storage Level: Disk Serialized 1x Replicated
Cached Partitions: 72
Total Partitions: 72
Memory Size: 42.7 GB
Disk Size: 73.3 GB
=> I understand : Memory Size: 42.7 GB + Disk Size: 73.3 GB = 110 G !
=> So my 6 G file has transformed on 37 G and then on 110 G ???
But i try to understand why is there still some memory left on my executor, and I go to the "err" dump of one, and I see :
18/02/08 11:04:08 INFO MemoryStore: Will not store rdd_50_46
18/02/08 11:04:09 WARN MemoryStore: Not enough space to cache rdd_50_46 in memory! (computed 1134.1 MB so far)
18/02/08 11:04:09 INFO MemoryStore: Memory use = 1641.6 KB (blocks) + 7.7 GB (scratch space shared across 6 tasks(s)) = 7.7 GB. Storage limit = 7.8 GB.
18/02/08 11:04:09 WARN BlockManager: Persisting block rdd_50_46 to disk instead.
And Here I see that the executor want to cache a 1641.6 KB block (only 1Mo !) and I can't because there is a ["scratch space"] of 7.7 Gb "shared across 6 tasks".
=> What is a "scratch space" ? ?
=> The 6 tasks => comes from the parallelism of 48 / 12 = 6
And then I come back to the app information, and I see that the count that lasted 48 min read only 37 Gb of data ! (The 48 min are clearly used to cache the data too)
When I do a count on the cached dataframe I have a 116G input read :
And at the end of the day, the time saved by the cached count is not that impressive, here are 3 duration :
4.8 ' : count on cached df
48' : count while caching
5.8' : count on not cached df (read directly from hdfs)
So why is it so ?
Because the cached df is not that much cached :
Meaning more or less 40 Gb in memory and 60 Gb on disk.
I am surprised because at 15G / executor * 12 slaves => 180 Gb of memory, and I can cache only 40 Gb ... But in fact I remember that the memory is splitted :
30% for spark
54% for storage
16% for shuffle
So I understand that I do have 54% * 15G for storage, ie 8.1 G, meaning that on my 180 Gb, I only have 97 Gb for storage. Why do I have 90 - 40 = 50 G not used then ?
Oups... This is a long post !
Plenty of questions... Sorry...
Any one know how many bytes occupy per file in namenode of Hdfs?
I want to estimate how many files can store in single namenode of 32G memory.
Each file or directory or block occupies about 150 bytes in the namenode memory. [1] So a cluster with a namenode with 32G RAM can support a maximum of (assuming namenode is the bottleneck) about 38 million files. (Each file will also take up a block, so each file takes 300 bytes in effect. I am also assuming 3x replication. So each file takes up 900 bytes)
In practice however, the number will be much lesser because all of the 32G will not be available to the namenode for keeping the mapping. You can increase it by allocating more heap space to the namenode in that machine.
Replication also effects this to a lesser degree. Each additional replica adds about 16 bytes to the memory requirement. [2]
[1] https://blog.cloudera.com/small-files-big-foils-addressing-the-associated-metadata-and-application-challenges/
[2] http://search-hadoop.com/c/HDFS:/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockInfo.java%7C%7CBlockInfo
Cloudera recommends 1 GB of NameNode heap space per million blocks. 1 GB for every million files is less conservative but should work too.
Also you don't need to multiply by a replication factor, an accepted answer is wrong.
Using the default block size of 128 MB, a file of 192 MB is split into two block files, one 128 MB file and one 64 MB file. On the NameNode, namespace objects are measured by the number of files and blocks. The same 192 MB file is represented by three namespace objects (1 file inode + 2 blocks) and consumes approximately 450 bytes of memory.
One data file of 128 MB is represented by two namespace objects on the NameNode (1 file inode + 1 block) and consumes approximately 300 bytes of memory. By contrast, 128 files of 1 MB each are represented by 256 namespace objects (128 file inodes + 128 blocks) and consume approximately 38,400 bytes.
Replication affects disk space but not memory consumption. Replication changes the amount of storage required for each block but not the number of blocks. If one block file on a DataNode, represented by one block on the NameNode, is replicated three times, the number of block files is tripled but not the number of blocks that represent them.
Examples:
1 x 1024 MB file
1 file inode
8 blocks (1024 MB / 128 MB)
Total = 9 objects * 150 bytes = 1,350 bytes of heap memory
8 x 128 MB files
8 file inodes
8 blocks
Total = 16 objects * 150 bytes = 2,400 bytes of heap memory
1,024 x 1 MB files
1,024 file inodes
1,024 blocks
Total = 2,048 objects * 150 bytes = 307,200 bytes of heap memory
Even more examples article in the origin article from cloudera.
(Each file metadata = 150bytes) + (block metadata for the file=150bytes)=300bytes
so 1million files each with 1 block will consume=300*1000000=300000000bytes
=300MB for replication factor of 1. with replication factor of 3 it requires 900MB.
So as thumb rule for every 1GB you can store 1million files.
We have the following stats on single node cassandra on Amazon EC2/Rightscale m1.large instance with 2 ephemeral disks with raid0. (7.6 GB Total Memory)
4 GB RAM is allocated to cassandra Heap, 800MB is Heap NEW size.
following stats are from OpsCenter community 2.0
Read Requests 285 to 340 per second
Write Requests 257 to 720 per second
OS Load 15.15 to 17.15
Write Request Latency 293 to 685 micros
OS Sent Network Traffic 18 MB to 30 MB per second
OS Recieved Network Traffic 22 MB to 34 MB per second
OS Disk Queue Size 23 to 26 requests
Read Requests Pending 8 to 20
Read Request Latency 69140 to 92885 micros
OS Disk latency 37 to 42 ms
OS Disk Throughput 12 to 14 Mb per second
Disk IOPs Reads 600 to 740 per second
Disk IOPs Writes 2 to 7 per second
IOWait 60 to 70 % CPU avg
Idle 24 to 30 % CPU avg
Rowcache is disabled.
Are the above stats are satisfying with the provided configuration....OR how could we tweak it more to get less IOWait..........because we think that we are experiencing lots of IOWait.....how could we tweak it to get the best.
Read Requests are mixed.........some are from one super column family and one standard having more than million keys......and varying no. of super columns max 14 with varying no. of subcolumns from 1 to 10000 and varying no. of columns max 14 in standard column family...............subcolumns are very thin in nature with 0 bytes value....8 bytes for name.
Process is removing the data from super column family and writing the processed data on standard one.
Would EBS Disks work better....on Amazon EC2
I'm not positive whether you can tweak your config easily to get more disk performance, but using Snappy compression could help a good deal in making your app need to read less overall. It may also help to use the new composite key layout instead of supercolumns.
One thing I can say for sure: EBS will NOT work better. Stay away from that at all costs if you care about latency.