Return-Path: X-Original-To: apmail-hadoop-common-user-archive@www.apache.org Delivered-To: apmail-hadoop-common-user-archive@www.apache.org Received: from mail.apache.org (hermes.apache.org [140.211.11.3]) by minotaur.apache.org (Postfix) with SMTP id F038510F8F for ; Sat, 15 Jun 2013 15:01:27 +0000 (UTC) Received: (qmail 96505 invoked by uid 500); 15 Jun 2013 15:01:22 -0000 Delivered-To: apmail-hadoop-common-user-archive@hadoop.apache.org Received: (qmail 96374 invoked by uid 500); 15 Jun 2013 15:01:22 -0000 Mailing-List: contact user-help@hadoop.apache.org; run by ezmlm Precedence: bulk List-Help: List-Unsubscribe: List-Post: List-Id: Reply-To: user@hadoop.apache.org Delivered-To: mailing list user@hadoop.apache.org Received: (qmail 96363 invoked by uid 99); 15 Jun 2013 15:01:22 -0000 Received: from athena.apache.org (HELO athena.apache.org) (140.211.11.136) by apache.org (qpsmtpd/0.29) with ESMTP; Sat, 15 Jun 2013 15:01:22 +0000 X-ASF-Spam-Status: No, hits=1.5 required=5.0 tests=HTML_MESSAGE,RCVD_IN_DNSWL_LOW,SPF_PASS,WEIRD_PORT X-Spam-Check-By: apache.org Received-SPF: pass (athena.apache.org: domain of harsh@cloudera.com designates 209.85.160.50 as permitted sender) Received: from [209.85.160.50] (HELO mail-pb0-f50.google.com) (209.85.160.50) by apache.org (qpsmtpd/0.29) with ESMTP; Sat, 15 Jun 2013 15:01:18 +0000 Received: by mail-pb0-f50.google.com with SMTP id wz7so1422092pbc.23 for ; Sat, 15 Jun 2013 08:00:57 -0700 (PDT) X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=google.com; s=20120113; h=mime-version:in-reply-to:references:from:date:message-id:subject:to :content-type:x-gm-message-state; bh=gozkVbDGQoyqrX1u8YuUn+9KAOTna5BcPaEhOuIWbAI=; b=pEi9ZuZ0pWHRToJ3QjhE32ifczKPsi31n1uMxfuxO56q9zKTEBcd4QMA8vDqcOepZ7 Pcfa7OZTrX8wuNc3zJSZvWeDb3L9iXIk5TZHX4PQG8qBSAmiw6iH7kABy1Yz3yTWTcrX uk289QSecgV03bmMjHu+xruMYcq3X5IU5fSR2yMp6mIl3qHSQrzlGsmxyAaDZvSZNnz1 cShY/kA+qRxJcCTKmMcMeAdPOFXJWFYJzqqNATpOdOiNlGihzYRFOjclzNUTNoYlwxMo 2XzfN8/g9scEUZFg8AxjxppXs3cV1Qvm7jpGPmq8iP8yv+KdQV6/OYekCrHuesbuCJm8 CYuA== X-Received: by 10.66.182.166 with SMTP id ef6mr6787352pac.35.1371308457654; Sat, 15 Jun 2013 08:00:57 -0700 (PDT) MIME-Version: 1.0 Received: by 10.70.75.168 with HTTP; Sat, 15 Jun 2013 08:00:37 -0700 (PDT) In-Reply-To: References: From: Harsh J Date: Sat, 15 Jun 2013 20:30:37 +0530 Message-ID: Subject: Re: Application errors with one disk on datanode getting filled up to 100% To: "" Content-Type: multipart/alternative; boundary=047d7bd6ba887f10c304df32a198 X-Gm-Message-State: ALoCoQmsNfX3NFJEHFxyANYd22Pp+/PA9cJjoQl2r5c7fCoa1x12rwBXPlAf5qUATNu2Tv8Zn3CY X-Virus-Checked: Checked by ClamAV on apache.org --047d7bd6ba887f10c304df32a198 Content-Type: text/plain; charset=ISO-8859-1 Sandeep/Mayank, If you take a look at the volume selection parts of the code, you can notice it is simply round robin. There's no way we continuously may select the same disk, unless the disk is deselected for errors (tolerated) or space (due to lack or reservation). Its better to monitor for a pattern and look for a misconfiguration, rather than suspect a bug and also accept the behavior. Rahul, The current HDFS version received a better inter-disk balancing code that I've seen in use already. See https://issues.apache.org/jira/browse/HDFS-1804 for more info. On Fri, Jun 14, 2013 at 4:45 PM, Sandeep L wrote: > Rahul, > > In general this issue happens some times in Hadoop. There is no exact > reason for this. > To mitigate this you need to run balancer in regular intervals. > > Thanks, > Sandeep. > > ------------------------------ > Date: Fri, 14 Jun 2013 16:39:02 +0530 > Subject: Re: Application errors with one disk on datanode getting filled > up to 100% > From: mail2mayank@gmail.com > To: user@hadoop.apache.org > > > No, as of this moment we've no ideas about the reasons for that behavior. > > > On Fri, Jun 14, 2013 at 4:04 PM, Rahul Bhattacharjee < > rahul.rec.dgp@gmail.com> wrote: > > Thanks Mayank, Any clue on why was only one disk was getting all writes. > > Rahul > > > On Thu, Jun 13, 2013 at 11:47 AM, Mayank wrote: > > So we did a manual rebalance (followed instructions at: > http://wiki.apache.org/hadoop/FAQ#On_an_individual_data_node.2C_how_do_you_balance_the_blocks_on_the_disk.3F) > and also reserved 30 GB of space for non dfs usage via > dfs.datanode.du.reserved and restarted our apps. > > Things have been going fine till now. > > Keeping fingers crossed :) > > > On Wed, Jun 12, 2013 at 12:58 PM, Rahul Bhattacharjee < > rahul.rec.dgp@gmail.com> wrote: > > I have a few points to make , these may not be very helpful for the said > problem. > > +All data nodes are bad exception is kind of not pointing to the problem > related to disk space full. > +hadoop.tmp.dir acts as base location of other hadoop related properties , > not sure if any particular directory is created specifically. > +Only one disk getting filled looks strange.The other disk are part while > formatting the NN. > > Would be interesting to know the reason for this. > Please keep posted. > > Thanks, > Rahul > > > On Mon, Jun 10, 2013 at 3:39 PM, Nitin Pawar wrote: > > From the snapshot, you got around 3TB for writing data. > > Can you check individual datanode's storage health. > As you said you got 80 servers writing parallely to hdfs, I am not sure > can that be an issue. > As suggested in past threads, you can do a rebalance of the blocks but > that will take some time to finish and will not solve your issue right > away. > > You can wait for others to reply. I am sure there will be far better > solutions from experts for this. > > > On Mon, Jun 10, 2013 at 3:18 PM, Mayank wrote: > > No it's not a map-reduce job. We've a java app running on around 80 > machines which writes to hdfs. The error that I'd mentioned is being thrown > by the application and yes we've replication factor set to 3 and following > is status of hdfs: > > Configured Capacity : 16.15 TB DFS Used : 11.84 TB Non DFS Used : 872.66 > GB DFS Remaining : 3.46 TB DFS Used% : 73.3 % DFS Remaining% : 21.42 % Live > Nodes :10 Dead > Nodes > : 0 Decommissioning Nodes > : 0 Number of Under-Replicated Blocks : 0 > > > On Mon, Jun 10, 2013 at 3:11 PM, Nitin Pawar wrote: > > when you say application errors out .. does that mean your mapreduce job > is erroring? In that case apart from hdfs space you will need to look at > mapred tmp directory space as well. > > you got 400GB * 4 * 10 = 16TB of disk and lets assume that you have a > replication factor of 3 so at max you will have datasize of 5TB with you. > I am also assuming you are not scheduling your program to run on entire > 5TB with just 10 nodes. > > i suspect your clusters mapred tmp space is getting filled in while the > job is running. > > > > > > On Mon, Jun 10, 2013 at 3:06 PM, Mayank wrote: > > We are running a hadoop cluster with 10 datanodes and a namenode. Each > datanode is setup with 4 disks (/data1, /data2, /data3, /data4), which each > disk having a capacity 414GB. > > > hdfs-site.xml has following property set: > > > dfs.data.dir > > /data1/hadoopfs,/data2/hadoopfs,/data3/hadoopfs,/data4/hadoopfs > Data dirs for DFS. > > > Now we are facing a issue where in we find /data1 getting filled up > quickly and many a times we see it's usage running at 100% with just few > megabytes of free space. This issue is visible on 7 out of 10 datanodes at > present. > > We've some java applications which are writing to hdfs and many a times we > are seeing foloowing errors in our application logs: > > > > java.io.IOException: All datanodes xxx.xxx.xxx.xxx:50010 are bad. Aborting... > at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.processDatanodeError(DFSClient.java:3093) > at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.access$2200(DFSClient.java:2586) > at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream$DataStreamer.run(DFSClient.java:2790) > > > > I went through some old discussions and looks like manual rebalancing is > what is required in this case and we should also have > dfs.datanode.du.reserved set up. > > However I'd like to understand if this issue, with one disk getting filled > up to 100% can result into the issue which we are seeing in our > application. > > Also, are there any other peformance implications due to some of the disks > running at 100% usage on a datanode. > -- > Mayank Joshi > > Skype: mail2mayank > Mb.: +91 8690625808 > > Blog: http://www.techynfreesouls.co.nr > PhotoStream: http://picasaweb.google.com/mail2mayank > > Today is tommorrow I was so worried about yesterday ... > > > > > -- > Nitin Pawar > > > > > -- > Mayank Joshi > > Skype: mail2mayank > Mb.: +91 8690625808 > > Blog: http://www.techynfreesouls.co.nr > PhotoStream: http://picasaweb.google.com/mail2mayank > > Today is tommorrow I was so worried about yesterday ... > > > > > -- > Nitin Pawar > > > > > > -- > Mayank Joshi > > Skype: mail2mayank > Mb.: +91 8690625808 > > Blog: http://www.techynfreesouls.co.nr > PhotoStream: http://picasaweb.google.com/mail2mayank > > Today is tommorrow I was so worried about yesterday ... > > > > > > -- > Mayank Joshi > > Skype: mail2mayank > Mb.: +91 8690625808 > > Blog: http://www.techynfreesouls.co.nr > PhotoStream: http://picasaweb.google.com/mail2mayank > > Today is tommorrow I was so worried about yesterday ... > -- Harsh J --047d7bd6ba887f10c304df32a198 Content-Type: text/html; charset=ISO-8859-1 Content-Transfer-Encoding: quoted-printable
Sandeep/Mayank,

If you take a look at t= he volume selection parts of the code, you can notice it is simply round ro= bin. There's no way we continuously may select the same disk, unless th= e disk is deselected for errors (tolerated) or space (due to lack or reserv= ation). Its better to monitor for a pattern and look for a misconfiguration= , rather than suspect a bug and also accept the behavior.

Rahul,

The c= urrent HDFS version received a better inter-disk balancing code that I'= ve seen in use already. See=A0https://issues.apache.org/jira/browse/HDFS-1804 for more= info.


On Fri,= Jun 14, 2013 at 4:45 PM, Sandeep L <sandeepvreddy@outlook.com= > wrote:
Rahul,

In general this issue happe= ns some times in Hadoop. There is no exact reason for this.
To mi= tigate this you need to run balancer in regular intervals.

Thanks,
Sandeep.

Date: Fri, 14 = Jun 2013 16:39:02 +0530
Subject: Re: Application errors with one disk on= datanode getting filled up to 100%
From: mail2mayank@gmail.com
To: user@hadoop= .apache.org


No, as of th= is moment we've no ideas about the reasons for that behavior.


On Fri, Jun 14, 2013 at 4:04 PM, Rahul Bhattacharjee <r= ahul.rec.dgp@gmail.com> wrote:
Thanks Mayank, Any clue o= n why was only one disk was getting all writes.

Rahul


On Thu, Jun = 13, 2013 at 11:47 AM, Mayank <mail2mayank@gmail.com> wro= te:
So we did a manual rebalance (followed instructions at: http://wiki.apache.or= g/hadoop/FAQ#On_an_individual_data_node.2C_how_do_you_balance_the_blocks_on= _the_disk.3F) and also reserved 30 GB of space for non dfs usage via df= s.datanode.du.reserved and restarted our apps.

Things have been going fine till now.

Keeping fingers crossed := )


On Wed, Jun 12, 2013 at 12:58 PM, Ra= hul Bhattacharjee <rahul.rec.dgp@gmail.com> wrote:
I have a few points to ma= ke , these may not be very helpful for the said problem.

+All data nodes are bad exception is kind of not pointing to the problem re= lated to disk space full.
+hadoop.tmp.dir acts as base location of other hadoop related properties ,= not sure if any particular directory is created specifically.
+Only one disk getting fille= d looks strange.The other disk are part while formatting the NN.

Would be interesting to know the reason for this.
Please keep posted.

Thanks,
Rahul


On Mon, Jun 10,= 2013 at 3:39 PM, Nitin Pawar <nitinpawar432@gmail.com> wrote:
From the snapshot, you got around 3TB for writing data.=A0
Can you check individual datanode's storage health.=A0
As you said you got 80 servers writing parallely to hdfs, I am not sur= e can that be an issue.=A0
As suggested in past threads, you can do a rebalance of the blocks but= that will take some time to finish and will not solve your issue right awa= y.=A0

You can wait for others to reply. I am sure = there will be far better solutions from experts for this.=A0


On Mon, Jun 10, 2013 at 3:18 PM, Mayank <= span dir=3D"ltr"><mail2mayank@gmail.com> wrote:
No it's not a map-reduce job. We've a java app running on = around 80 machines which writes to hdfs. The error that I'd mentioned i= s being thrown by the application and yes we've replication factor set = to 3 and following is status of hdfs:

21.42 %<= /table>


On Mon, Jun 10, 2013 at 3:11 P= M, Nitin Pawar <nitinpawar432@gmail.com> wrote:
when you say application errors out .. does that mean your mapredu= ce job is erroring? In that case apart from hdfs space you will need to loo= k at mapred tmp directory space as well.=A0

you got 400GB * 4 * 10 =3D 16TB of disk and lets assume that you have a rep= lication factor of 3 so at max you will have datasize of 5TB with you.=A0
I am also assuming you are not scheduling your program to run on e= ntire 5TB with just 10 nodes.=A0

i suspect your clusters mapred tmp space is getting fil= led in while the job is running.=A0





On Mon, Jun 10, 2013 at 3:06 PM, Mayank <mail2mayank@gmail.co= m> wrote:
We are running a hadoop cluster with 10 datanodes and= a namenode. Each datanode is setup with 4 disks (/data1, /data2, /data3, /= data4), which each disk having a capacity 414GB.


hdfs= -site.xml has following property set:

<property>
=A0=A0=A0=A0=A0=A0=A0 <name>dfs.data.dir</= name>
=A0=A0=A0=A0=A0=A0=A0 <value>/data1/hadoopfs,/data2/hadoo= pfs,/data3/hadoopfs,/data4/hadoopfs</value>
=A0=A0=A0=A0=A0=A0=A0 = <description>Data dirs for DFS.</description>
</property>

Now we are facing a iss= ue where in we find /data1 getting filled up quickly and many a times we se= e it's usage running at 100% with just few megabytes of free space. Thi= s issue is visible on 7 out of 10 datanodes at present.

We've some java applications which are writing to hdfs a= nd many a times we are seeing foloowing errors in our application logs:
= 


java.io.IOException: All datanodes xxx.xxx.xxx.xxx:50010 are bad. Aborting...
	at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.processDatanodeError(D=
FSClient.java:3093)
	at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream.access$2200(DFSClient.=
java:2586)
	at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream$DataStreamer.run(DFSCl=
ient.java:2790)
I went through some old discussions and looks like manual rebalan= cing is what is required in this case and we should also have dfs.datanode.= du.reserved set up.

However I'd like to understand if= this issue, with one disk getting filled up to 100% can result into the is= sue which we are seeing in our application.

Also, are there any other peformance implications due to some of the di= sks running at 100% usage on a datanode.
<= /font>
--
Mayank Joshi

Skype: mail2mayank
Mb.:=A0 +91 8690625808=

Blog: http://www.techynfreesouls.co.nr
PhotoStream: http://picasaweb.google.com/mail2mayank

Today is tommorr= ow I was so worried about yesterday ...



<= font color=3D"#888888">--
Nitin Pawar



--
Mayank Joshi

Sky= pe: mail2mayank
Mb.:=A0 +91 8690625808

Blog: http://www.techynfreeso= uls.co.nr
PhotoStream: http://picasaweb.google.com/mail2mayank

Today is tommorr= ow I was so worried about yesterday ...



<= /div>--
Nitin Pawar




--
Mayank Josh= i

Skype: mail2mayank
Mb.:=A0 +91 8690625808

Blog: = http://www.= techynfreesouls.co.nr
PhotoStream: http://picasaweb.google.com/mail2mayank

Today is tommorr= ow I was so worried about yesterday ...



--
Mayank Josh= i

Skype: mail2mayank
Mb.:=A0 +91 8690625808

Blog: <= a href=3D"http://www.techynfreesouls.co.nr/" target=3D"_blank">http://www.t= echynfreesouls.co.nr
PhotoStream: http://picasaweb.google.com/mail2mayank

Today is tommorr= ow I was so worried about yesterday ...


--
Harsh J --047d7bd6ba887f10c304df32a198--
Configured Capacity : 16.15 TB=
DFS Used : 11.84 TB
Non DFS Used : 872.66 GB
DFS Rem= aining : 3.46 TB
DFS Used% : 73.3 %
DFS Remaining% :
Live Nodes : 10
Dead Nod= es : 0
Decommissioning Node= s : 0
Number of Under-Replicated Blocks : 0