Yes, I have done all the suggestion of the
http://wiki.apache.org/hadoop/PerformanceTuning.
I just restarted the hbase cluster and recreated the table, the data
insertion looks fine for now and
I am getting about 1k record/second . I consider that to be reasonable
giving that my record is about
10k bytes per record. but this is the beginning of the writing and I notice
that when the table is small,
the hbase works fine. when there are lots of records in the table already,
problem begin to happen.
I will report back and see how it goes after some more time.
Jimmy.
--------------------------------------------------
From: "Ryan Rawson" <[email protected]>
Sent: Wednesday, June 09, 2010 5:20 PM
To: <[email protected]>
Subject: Re: ideas to improve throughput of the base writting
I am not familiar with that exception, I have not seen of it before...
perhaps someone else has?
And my 200k rows/sec is over 19 machines. It is the average over many
hours. My calculation of row size might not match how much data was
flowing to disk, but I think it isn't too far off.
Unfortunately comparing raw disk speed in a trivial benchmark (such as
hdparm -t is) doesn't tell us how absolute speed of HBase must
perform. This is because HBase does much more work than a raw disk
write benchmark -- doing so to maintain structure and sorting. We can
say that 'faster disks = faster HBase performance'.
From the log lines you have pasted it sounds like the regionserver's
flush ability is not keeping up with your rate of data input. How big
are your records? What is your target input speed? Have you done
anything on this page:
http://wiki.apache.org/hadoop/PerformanceTuning
On Wed, Jun 9, 2010 at 4:58 PM, Jinsong Hu <[email protected]> wrote:
My hardware has 2 disks. I did a file copy on the machine and found that
I
can get 300 mbyte/second.
At this time, I see my insertion is less than 1k/second. my row size is .
in
terms of disk writing. my record
insertion rate is far less than the hardware limit. my row size is about
10K byte
if in your i7-based server, you are doing 200k row/sec, each row is 200
byte, then you are doing 40M byte/sec.
in my case, if it behaves normally, I can get 100 row/sec * 10K byte =1M
/sec.
that is far from the disk speed. occasionally I can see 1k row/second.
which
is more reasonable in my case,
but I rarely get that performance.
even worse, with the change done, now I have seem lots of compaction
failure:
2010-06-09 23:40:55,117 ERROR
org.apache.hadoop.hbase.regionserver.CompactSplitT
hread: Compaction failed for region Spam_MsgEventTable,2010-06-09
20:05:20\x0905
860d4bf1cb268ef69391cf97de9f64,1276121160527
java.lang.RuntimeException: java.io.IOException: Could not find target
position
65588
at
org.apache.hadoop.hbase.regionserver.StoreFileScanner.next(StoreFileS
canner.java:61)
at
org.apache.hadoop.hbase.regionserver.KeyValueHeap.next(KeyValueHeap.j
ava:79)
at
org.apache.hadoop.hbase.regionserver.MinorCompactingStoreScanner.next
(MinorCompactingStoreScanner.java:96)
at
org.apache.hadoop.hbase.regionserver.Store.compact(Store.java:920)
at
org.apache.hadoop.hbase.regionserver.Store.compact(Store.java:764)
at
org.apache.hadoop.hbase.regionserver.HRegion.compactStores(HRegion.ja
va:832)
at
org.apache.hadoop.hbase.regionserver.HRegion.compactStores(HRegion.ja
va:785)
at
org.apache.hadoop.hbase.regionserver.CompactSplitThread.run(CompactSp
litThread.java:93)
Caused by: java.io.IOException: Could not find target position 65588
at
org.apache.hadoop.hdfs.DFSClient$DFSInputStream.fetchBlockAt(DFSClien
t.java:1556)
at
org.apache.hadoop.hdfs.DFSClient$DFSInputStream.blockSeekTo(DFSClient
.java:1666)
at
org.apache.hadoop.hdfs.DFSClient$DFSInputStream.read(DFSClient.java:1
780)
at java.io.DataInputStream.read(DataInputStream.java:132)
at
org.apache.hadoop.hbase.io.hfile.BoundedRangeFileInputStream.read(Bou
ndedRangeFileInputStream.java:105)
at org.apache.hadoop.io.IOUtils.readFully(IOUtils.java:100)
at
org.apache.hadoop.hbase.io.hfile.HFile$Reader.decompress(HFile.java:1
018)
at
org.apache.hadoop.hbase.io.hfile.HFile$Reader.readBlock(HFile.java:96
6)
at
org.apache.hadoop.hbase.io.hfile.HFile$Reader$Scanner.next(HFile.java
:1159)
at
org.apache.hadoop.hbase.regionserver.StoreFileScanner.next(StoreFileS
canner.java:58)
... 7 more
I can't stop this unless I restarted the regionserver. After restart I
truncate the table, and when I list the table again in shell,
it appears 2 times. now I can't even disable the table and drop it.
I will restart the whole hbase cluster and report the progress.
Jimmy/
--------------------------------------------------
From: "Ryan Rawson" <[email protected]>
Sent: Wednesday, June 09, 2010 4:16 PM
To: <[email protected]>
Subject: Re: ideas to improve throughput of the base writting
Hey,
Sounds like you are hitting limits of your hardware... I dont think
you mentioned the hardware spec you are running in this thread...
What you are seeing is essentially the limits of HDFS's ability to
take writes. The errors might be due to various HDFS setup problems
(eg: xceiver count, file handle count, all outlined in various HBase
"startup" docs)... But the overall performance might be limited by
your hardware.
For example, I use i7-based servers with 4 disks. This gives a
reasonable IO bandwidth, and can cope with high rates of inserts (upto
100-200k rows/sec (each row is ~ 100-300 bytes). If you are running a
1 or 2 disk system it is possible you are hitting limits of what your
hardware can do.
Also note that the write-pipeline performance is ultimately defined in
bytes/sec, not just 'rows/sec'... thus my rows were small, and if
yours are big then you might be hitting a lower 'row/sec' limit even
though the amount of bytes you are writing is higher than what i might
have been doing.
On Wed, Jun 9, 2010 at 3:59 PM, Jinsong Hu <[email protected]>
wrote:
I still get lots of repetition of
2010-06-09 22:54:38,428 WARN
org.apache.hadoop.hbase.regionserver.MemStoreFlushe
r: Region Spam_MsgEventTable,2010-06-09
20:05:20\x0905860d4bf1cb268ef69391cf97de
9f64,1276121160527 has too many store files, putting it back at the end
of
the f
lush queue.
2010-06-09 22:54:38,428 DEBUG
org.apache.hadoop.hbase.regionserver.CompactSplitT
hread: Compaction requested for region Spam_MsgEventTable,2010-06-09
20:05:20\x0
905860d4bf1cb268ef69391cf97de9f64,1276121160527/1537478401 because:
regionserver
/10.110.8.88:60020.cacheFlusher
I also saw lots of
2010-06-09 22:50:12,527 INFO
org.apache.hadoop.hbase.regionserver.HRegion:
Block
ing updates for 'IPC Server handler 1 on 60020' on region
Spam_MsgEventTable,201
0-06-09 20:05:20\x0905860d4bf1cb268ef69391cf97de9f64,1276121160527:
memstore
siz
e 512.0m is >= than blocking 512.0m size
2010-06-09 22:50:12,598 INFO
org.apache.hadoop.hbase.regionserver.HRegion:
Block
ing updates for 'IPC Server handler 5 on 60020' on region
Spam_MsgEventTable,201
0-06-09 20:05:20\x0905860d4bf1cb268ef69391cf97de9f64,1276121160527:
memstore
siz
e 512.0m is >= than blocking 512.0m size
even with the changed config. the regionserver has 4G ram. what else
can
be
wrong ?
The insertion rate is still not good.
Jimmy.
--------------------------------------------------
From: "Jinsong Hu" <[email protected]>
Sent: Wednesday, June 09, 2010 1:59 PM
To: <[email protected]>
Subject: Re: ideas to improve throughput of the base writting
Thanks. I will make this change:
<property>
<name>hbase.hregion.memstore.block.multiplier</name>
<value>8</value>
</property>
<property>
<name>hbase.regionserver.msginterval</name>
<value>10000</value>
</property>
<property>
<name>hbase.hstore.compactionThreshold</name>
<value>6</value>
</property>
<property>
<name>hbase.hstore.blockingStoreFiles</name>
<value>18</value>
</property>
and see how it goes.
Jimmy.
--------------------------------------------------
From: "Ryan Rawson" <[email protected]>
Sent: Wednesday, June 09, 2010 1:49 PM
To: <[email protected]>
Subject: Re: ideas to improve throughput of the base writting
More background here... you are running into a situation where the
regionserver cannot flush fast enough and the size of the region's
memstore has climbed too high and thus you get that error message.
HBase attempts to protect itself by holding up clients (thus causing
the low performance you see). By expanding how big a memstore can
get
during times of stress you can improve performance, at the cost of
memory usage. That is what that setting is about.
As for the 1.5 minute setting, that is the maximal amount of time a
handler thread will block for. You shouldn't need to tweak that
value, and reducing it could cause issues.
Now, as for compacting, HBase will compact small files into larger
files, and on a massive upload you can expect to see this happen
constantly, thus tying up 1 cpu worth on your regionserver. You
could
potentially reduce that by increasing the value:
<property>
<name>hbase.hstore.compactionThreshold</name>
<value>3</value>
the value is interpreted as "if there are more than 3 files for a
region then run the compaction check". By raising this limit you can
accumulate more files before compacting them, thus reducing the
frequency of compactions but also potentially increasing the
performance of reads (more files to read = more seeks = slower). I'd
consider setting it to 5-7 or so in concert with setting
"hbase.hstore.blockingStoreFiles" to a value at least 2x that.
All of these settings increase the amount of ram your regionserver
may
need, so you will want to ensure you have at least 4000m of ram set
in
hbase-env.sh. This is why they are set so conservatively in the
default shipping config.
These are the 3 important settings that control how often compactions
occur and how RPC threads get blocked. Try tweaking all of them and
let me know if you are doing better.
-ryan
On Wed, Jun 9, 2010 at 1:37 PM, Ryan Rawson <[email protected]>
wrote:
you also want this config:
<property>
<name>hbase.hregion.memstore.block.multiplier</name>
<value>8</value>
</property>
that should hopefully clear things up.
-ryan
On Wed, Jun 9, 2010 at 1:34 PM, Jinsong Hu <[email protected]>
wrote:
I checked the log, there are lots of
e 128.1m is >= than blocking 128.0m size
2010-06-09 17:26:36,736 INFO
org.apache.hadoop.hbase.regionserver.HRegion:
Block
ing updates for 'IPC Server handler 8 on 60020' on region
Spam_MsgEventTable,201
0-06-09 05:25:32\x09c873847edf6e5390477494956ec04729,1276104002262:
memstore
siz
e 128.1m is >= than blocking 128.0m size
then after that there are lots of
2010-06-09 17:26:36,800 DEBUG
org.apache.hadoop.hbase.regionserver.Store:
Added
hdfs://namenodes1.cloud.ppops.net:8020/hbase/Spam_MsgEventTable/376337880/messag
e_compound_terms/7606939244559826252, entries=30869,
sequenceid=8350447892,
mems
ize=7.2m, filesize=3.4m to Spam_MsgEventTable,2010-06-09
05:25:32\x09c873847edf6
then lots of
2010-06-09 17:26:39,005 INFO
org.apache.hadoop.hbase.regionserver.HRegion:
Unblo
cking updates for region Spam_MsgEventTable,2010-06-09
05:25:32\x09c873847edf6e5
390477494956ec04729,1276104002262 'IPC Server handler 8 on 60020'
This cycle happens again and again in the log. What can I do in
this
case
to speed up writing ?
right now the writing speed is really slow. close to 4 rows/second
for
a
regionserver.
I checked the code and try to find out why there are so many store
files,
and I noticed each second
the regionserver reports to master, it calls the memstore flush and
write a
store file.
the parameter hbase.regionserver.msginterval default value is 1
second.
I am
thinking to change to 10 second.
can that help ? I am also thinking to change
hbase.hstore.blockingStoreFiles
to 1000. I noticed that there is a parameter
hbase.hstore.blockingWaitTime with default value of 1.5 minutes. as
long as
the 1.5 minutes is reached,
the compaction is executed. I am fine with running compaction every
1.5
minutes, but running compaction every second
and causing CPU consistently higher than 100% is not wanted.
Any suggestion what kind of parameters to change to improve my
writing
speed
?
Jimmy
--------------------------------------------------
From: "Ryan Rawson" <[email protected]>
Sent: Wednesday, June 09, 2010 1:01 PM
To: <[email protected]>
Subject: Re: ideas to improve throughput of the base writting
The log will say something like "blocking updates to..." when you
hit
a limit. That log you indicate is just the regionserver
attempting
to
compact a region, but shouldn't prevent updates.
what else does your logfile say? Search for the string (case
insensitive) "blocking updates"...
-ryan
On Wed, Jun 9, 2010 at 11:52 AM, Jinsong Hu
<[email protected]>
wrote:
I made this change
<property>
<name>hbase.hstore.blockingStoreFiles</name>
<value>15</value>
</property>
the system is still slow.
Here is the most recent value for the region :
stores=21, storefiles=186, storefileSizeMB=9681,
memstoreSizeMB=128,
storefileIndexSizeMB=12
And the same log still happens:
2010-06-09 18:36:40,577 WARN org.apache.h
adoop.hbase.regionserver.MemStoreFlusher: Region
SOME_ABCEventTable,2010-06-09 0
9:56:56\x093dc01b4d2c4872963717d80d8b5c74b1,1276107447570 has too
many
store
fil
es, putting it back at the end of the flush queue.
One idea that I have now is to further increase the
hbase.hstore.blockingStoreFiles to a very high
Number, such as 1000. What is the negative impact of this change
?
Jimmy
--------------------------------------------------
From: "Ryan Rawson" <[email protected]>
Sent: Monday, June 07, 2010 3:58 PM
To: <[email protected]>
Subject: Re: ideas to improve throughput of the base writting
Try setting this config value:
<property>
<name>hbase.hstore.blockingStoreFiles</name>
<value>15</value>
</property>
and see if that helps.
The thing about the 1 compact thread is the scarce resources
being
preserved in this case is cluster IO. People have had issues
with
compaction IO being too heavy.
in your case, this setting can let the regionserver build up
more
store files without pausing your import.
-ryan
On Mon, Jun 7, 2010 at 3:52 PM, Jinsong Hu
<[email protected]>
wrote:
Hi, There:
While saving lots of data to on hbase, I noticed that the
regionserver
CPU
went to more than 100%. examination shows that the hbase
CompactSplit
is
spending full time working on compacting/splitting hbase store
files.
The
machine I have is an 8 core machine. because there is only one
comact/split
thread in hbase, only one core is fully used.
I continue to submit map/reduce job to insert records to
hbase.
most
of
the time, the job runs very fast, around 1-5 minutes. But
occasionally,
it
can take 2 hours. That is very bad to me. I highly suspect that
the
occasional slow insertion is related to the
insufficient speed compactsplit thread.
I am thinking that I should parallize the compactsplit thread,
the
code
has
this : the for loop "for (Store store: stores.values()) " can
be
parallized via java 5's threadpool , thus multiple cores are
used
instead
only one core is used. I wonder if this will help to increase
the
throughput.
Somebody mentioned that I can increase the regionsize to that
I
don't
do
so
many compaction. Under heavy writing situation.
does anybody have experience showing it helps ?
Jimmy.
byte [] compactStores(final boolean majorCompaction)
throws IOException {
if (this.closing.get() || this.closed.get()) {
LOG.debug("Skipping compaction on " + this + " because
closing/closed");
return null;
}
splitsAndClosesLock.readLock().lock();
try {
byte [] splitRow = null;
if (this.closed.get()) {
return splitRow;
}
try {
