[Gluster-users] disperse volume file to subvolume mapping

Thu Apr 21 08:24:58 UTC 2016

Hi Serkan,

On 21/04/16 10:07, Serkan Çoban wrote:
>> I think the problem is in the temporary name that distcp gives to the file while it's being copied before renaming it to the real name. Do you know what is the structure of this name ?
> Distcp temporary file name format is:
> ".distcp.tmp.attempt_1460381790773_0248_m_000001_0" and the same
> temporary file name used by one map process. For example I see in the
> logs that one map copies files part-m-00031,part-m-00047,part-m-00063
> sequentially and they all use same temporary file name above. So no
> original file name appears in temporary file name.

This explains the problem. With the default options, DHT sends all files 
to the subvolume that should store a file named 'distcp.tmp'.

With this temporary name format, little can be done.

>
> I will check if we can modify distcp behaviour, or we have to write
> our mapreduce procedures instead of using distcp.
>
>> 2. define the option 'extra-hash-regex' to an expression that matches your temporary file names and returns the same name that will finally have. Depending on the differences between original and temporary file names, this option could be useless.
>> 3. set the option 'rsync-hash-regex' to 'none'. This will prevent the name conversion, so the files will be evenly distributed. However this will cause a lot of files placed in incorrect subvolumes, creating a lot of link files until a rebalance is executed.
>
> How can I set these options?

You can set gluster options using:

gluster volume set <volname> <option> <value>

for example:

gluster volume set v0 rsync-hash-regex none

Xavi

>
>
>
> On Thu, Apr 21, 2016 at 10:00 AM, Xavier Hernandez
> <xhernandez at datalab.es> wrote:
>> Hi Serkan,
>>
>> I think the problem is in the temporary name that distcp gives to the file
>> while it's being copied before renaming it to the real name. Do you know
>> what is the structure of this name ?
>>
>> DHT selects the subvolume (in this case the ec set) on which the file will
>> be stored based on the name of the file. This has a problem when a file is
>> being renamed, because this could change the subvolume where the file should
>> be found.
>>
>> DHT has a feature to avoid incorrect file placements when executing renames
>> for the rsync case. What it does is to check if the file matches the
>> following regular expression:
>>
>>      ^\.(.+)\.[^.]+$
>>
>> If a match is found, it only considers the part between parenthesis to
>> calculate the destination subvolume.
>>
>> This is useful for rsync because temporary file names are constructed in the
>> following way: suppose the original filename is 'test'. The temporary
>> filename while rsync is being executed is made by prepending a dot and
>> appending '.<random chars>': .test.712hd
>>
>> As you can see, the original name and the part of the name between
>> parenthesis that matches the regular expression are the same. This causes
>> that, after renaming the temporary file to its original filename, both files
>> will be considered to belong to the same subvolume by DHT.
>>
>> In your case it's very probable that distcp uses a temporary name like
>> '.part.<number>'. In this case the portion of the name used to select the
>> subvolume is always 'part'. This would explain why all files go to the same
>> subvolume. Once the file is renamed to another name, DHT realizes that it
>> should go to another subvolume. At this point it creates a link file (those
>> files with access rights = '---------T') in the correct subvolume but it
>> doesn't move it. As you can see, this kind of files are better balanced.
>>
>> To solve this problem you have three options:
>>
>> 1. change the temporary filename used by distcp to correctly match the
>> regular expression. I'm not sure if this can be configured, but if this is
>> possible, this is the best option.
>>
>> 2. define the option 'extra-hash-regex' to an expression that matches your
>> temporary file names and returns the same name that will finally have.
>> Depending on the differences between original and temporary file names, this
>> option could be useless.
>>
>> 3. set the option 'rsync-hash-regex' to 'none'. This will prevent the name
>> conversion, so the files will be evenly distributed. However this will cause
>> a lot of files placed in incorrect subvolumes, creating a lot of link files
>> until a rebalance is executed.
>>
>> Xavi
>>
>>
>> On 20/04/16 14:13, Serkan Çoban wrote:
>>>
>>> Here is the steps that I do in detail and relevant output from bricks:
>>>
>>> I am using below command for volume creation:
>>> gluster volume create v0 disperse 20 redundancy 4 \
>>> 1.1.1.{185..204}:/bricks/02 \
>>> 1.1.1.{205..224}:/bricks/02 \
>>> 1.1.1.{225..244}:/bricks/02 \
>>> 1.1.1.{185..204}:/bricks/03 \
>>> 1.1.1.{205..224}:/bricks/03 \
>>> 1.1.1.{225..244}:/bricks/03 \
>>> 1.1.1.{185..204}:/bricks/04 \
>>> 1.1.1.{205..224}:/bricks/04 \
>>> 1.1.1.{225..244}:/bricks/04 \
>>> 1.1.1.{185..204}:/bricks/05 \
>>> 1.1.1.{205..224}:/bricks/05 \
>>> 1.1.1.{225..244}:/bricks/05 \
>>> 1.1.1.{185..204}:/bricks/06 \
>>> 1.1.1.{205..224}:/bricks/06 \
>>> 1.1.1.{225..244}:/bricks/06 \
>>> 1.1.1.{185..204}:/bricks/07 \
>>> 1.1.1.{205..224}:/bricks/07 \
>>> 1.1.1.{225..244}:/bricks/07 \
>>> 1.1.1.{185..204}:/bricks/08 \
>>> 1.1.1.{205..224}:/bricks/08 \
>>> 1.1.1.{225..244}:/bricks/08 \
>>> 1.1.1.{185..204}:/bricks/09 \
>>> 1.1.1.{205..224}:/bricks/09 \
>>> 1.1.1.{225..244}:/bricks/09 \
>>> 1.1.1.{185..204}:/bricks/10 \
>>> 1.1.1.{205..224}:/bricks/10 \
>>> 1.1.1.{225..244}:/bricks/10 \
>>> 1.1.1.{185..204}:/bricks/11 \
>>> 1.1.1.{205..224}:/bricks/11 \
>>> 1.1.1.{225..244}:/bricks/11 \
>>> 1.1.1.{185..204}:/bricks/12 \
>>> 1.1.1.{205..224}:/bricks/12 \
>>> 1.1.1.{225..244}:/bricks/12 \
>>> 1.1.1.{185..204}:/bricks/13 \
>>> 1.1.1.{205..224}:/bricks/13 \
>>> 1.1.1.{225..244}:/bricks/13 \
>>> 1.1.1.{185..204}:/bricks/14 \
>>> 1.1.1.{205..224}:/bricks/14 \
>>> 1.1.1.{225..244}:/bricks/14 \
>>> 1.1.1.{185..204}:/bricks/15 \
>>> 1.1.1.{205..224}:/bricks/15 \
>>> 1.1.1.{225..244}:/bricks/15 \
>>> 1.1.1.{185..204}:/bricks/16 \
>>> 1.1.1.{205..224}:/bricks/16 \
>>> 1.1.1.{225..244}:/bricks/16 \
>>> 1.1.1.{185..204}:/bricks/17 \
>>> 1.1.1.{205..224}:/bricks/17 \
>>> 1.1.1.{225..244}:/bricks/17 \
>>> 1.1.1.{185..204}:/bricks/18 \
>>> 1.1.1.{205..224}:/bricks/18 \
>>> 1.1.1.{225..244}:/bricks/18 \
>>> 1.1.1.{185..204}:/bricks/19 \
>>> 1.1.1.{205..224}:/bricks/19 \
>>> 1.1.1.{225..244}:/bricks/19 \
>>> 1.1.1.{185..204}:/bricks/20 \
>>> 1.1.1.{205..224}:/bricks/20 \
>>> 1.1.1.{225..244}:/bricks/20 \
>>> 1.1.1.{185..204}:/bricks/21 \
>>> 1.1.1.{205..224}:/bricks/21 \
>>> 1.1.1.{225..244}:/bricks/21 \
>>> 1.1.1.{185..204}:/bricks/22 \
>>> 1.1.1.{205..224}:/bricks/22 \
>>> 1.1.1.{225..244}:/bricks/22 \
>>> 1.1.1.{185..204}:/bricks/23 \
>>> 1.1.1.{205..224}:/bricks/23 \
>>> 1.1.1.{225..244}:/bricks/23 \
>>> 1.1.1.{185..204}:/bricks/24 \
>>> 1.1.1.{205..224}:/bricks/24 \
>>> 1.1.1.{225..244}:/bricks/24 \
>>> 1.1.1.{185..204}:/bricks/25 \
>>> 1.1.1.{205..224}:/bricks/25 \
>>> 1.1.1.{225..244}:/bricks/25 \
>>> 1.1.1.{185..204}:/bricks/26 \
>>> 1.1.1.{205..224}:/bricks/26 \
>>> 1.1.1.{225..244}:/bricks/26 \
>>> 1.1.1.{185..204}:/bricks/27 \
>>> 1.1.1.{205..224}:/bricks/27 \
>>> 1.1.1.{225..244}:/bricks/27 force
>>>
>>> then I mount volume on 50 clients:
>>> mount -t glusterfs 1.1.1.185:/v0 /mnt/gluster
>>>
>>> then I make a directory from one of the clients and chmod it.
>>> mkdir /mnt/gluster/s1 && chmod 777 /mnt/gluster/s1
>>>
>>> then I start distcp on clients, there are 1059X8.8GB files in one folder
>>> and
>>> they will be copied to /mnt/gluster/s1 with 100 parallel which means 2
>>> copy jobs per client at same time.
>>> hadoop distcp -m 100 http://nn1:8020/path/to/teragen-10tb
>>> file:///mnt/gluster/s1
>>>
>>> After job finished here is the status of s1 directory from bricks:
>>> s1 directory is present in all 1560 brick.
>>> s1/teragen-10tb folder is present in all 1560 brick.
>>>
>>> full listing of files in bricks:
>>> https://www.dropbox.com/s/rbgdxmrtwz8oya8/teragen_list.zip?dl=0
>>>
>>> You can ignore the .crc files in the brick output above, they are
>>> checksum files...
>>>
>>> As you can see part-m-xxxx files written only some bricks in nodes
>>> 0205..0224
>>> All bricks have some files but they have zero size.
>>>
>>> I increase file descriptors to 65k so it is not the issue...
>>>
>>>
>>>
>>>
>>>
>>> On Wed, Apr 20, 2016 at 9:34 AM, Xavier Hernandez <xhernandez at datalab.es>
>>> wrote:
>>>>
>>>> Hi Serkan,
>>>>
>>>> On 19/04/16 15:16, Serkan Çoban wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>> I assume that gluster is used to store the intermediate files before
>>>>>>>> the reduce phase
>>>>>
>>>>>
>>>>> Nope, gluster is the destination for distcp command. hadoop distcp -m
>>>>> 50 http://nn1:8020/path/to/folder file:///mnt/gluster
>>>>> This run maps on datanodes which have /mnt/gluster mounted on all of
>>>>> them.
>>>>
>>>>
>>>>
>>>> I don't know hadoop, so I'm of little help here. However it seems that -m
>>>> 50
>>>> means to execute 50 copies in parallel. This means that even if the
>>>> distribution worked fine, at most 50 (much probably less) of the 78 ec
>>>> sets
>>>> would be used in parallel.
>>>>
>>>>>
>>>>>>>> This means that this is caused by some peculiarity of the mapreduce.
>>>>>
>>>>>
>>>>> Yes but how a client write 500 files to gluster mount and those file
>>>>> just written only to subset of subvolumes? I cannot use gluster as a
>>>>> backup cluster if I cannot write with distcp.
>>>>>
>>>>
>>>> All 500 files were created only on one of the 78 ec sets and the
>>>> remaining
>>>> 77 got empty ?
>>>>
>>>>>>>> You should look which files are created in each brick and how many
>>>>>>>> while the process is running.
>>>>>
>>>>>
>>>>> Files only created on nodes 185..204 or 205..224 or 225..244. Only on
>>>>> 20 nodes in each test.
>>>>
>>>>
>>>>
>>>> How many files there were in each brick ?
>>>>
>>>> Not sure if this can be related, but standard linux distributions have a
>>>> default limit of 1024 open file descriptors. Having a so big volume and
>>>> doing a massive copy, maybe this limit is affecting something ?
>>>>
>>>> Are there any error or warning messages in the mount or bricks logs ?
>>>>
>>>>
>>>> Xavi
>>>>
>>>>>
>>>>> On Tue, Apr 19, 2016 at 1:05 PM, Xavier Hernandez
>>>>> <xhernandez at datalab.es>
>>>>> wrote:
>>>>>>
>>>>>>
>>>>>> Hi Serkan,
>>>>>>
>>>>>> moved to gluster-users since this doesn't belong to devel list.
>>>>>>
>>>>>> On 19/04/16 11:24, Serkan Çoban wrote:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> I am copying 10.000 files to gluster volume using mapreduce on
>>>>>>> clients. Each map process took one file at a time and copy it to
>>>>>>> gluster volume.
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> I assume that gluster is used to store the intermediate files before
>>>>>> the
>>>>>> reduce phase.
>>>>>>
>>>>>>> My disperse volume consist of 78 subvolumes of 16+4 disk each. So If I
>>>>>>> copy >78 files parallel I expect each file goes to different subvolume
>>>>>>> right?
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> If you only copy 78 files, most probably you will get some subvolume
>>>>>> empty
>>>>>> and some other with more than one or two files. It's not an exact
>>>>>> distribution, it's a statistially balanced distribution: over time and
>>>>>> with
>>>>>> enough files, each brick will contain an amount of files in the same
>>>>>> order
>>>>>> of magnitude, but they won't have the *same* number of files.
>>>>>>
>>>>>>> In my tests during tests with fio I can see every file goes to
>>>>>>> different subvolume, but when I start mapreduce process from clients
>>>>>>> only 78/3=26 subvolumes used for writing files.
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> This means that this is caused by some peculiarity of the mapreduce.
>>>>>>
>>>>>>> I see that clearly from network traffic. Mapreduce on client side can
>>>>>>> be run multi thread. I tested with 1-5-10 threads on each client but
>>>>>>> every time only 26 subvolumes used.
>>>>>>> How can I debug the issue further?
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> You should look which files are created in each brick and how many
>>>>>> while
>>>>>> the
>>>>>> process is running.
>>>>>>
>>>>>> Xavi
>>>>>>
>>>>>>
>>>>>>>
>>>>>>> On Tue, Apr 19, 2016 at 11:22 AM, Xavier Hernandez
>>>>>>> <xhernandez at datalab.es> wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> Hi Serkan,
>>>>>>>>
>>>>>>>> On 19/04/16 09:18, Serkan Çoban wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Hi, I just reinstalled fresh 3.7.11 and I am seeing the same
>>>>>>>>> behavior.
>>>>>>>>> 50 clients copying part-0-xxxx named files using mapreduce to
>>>>>>>>> gluster
>>>>>>>>> using one thread per server and they are using only 20 servers out
>>>>>>>>> of
>>>>>>>>> 60. On the other hand fio tests use all the servers. Anything I can
>>>>>>>>> do
>>>>>>>>> to solve the issue?
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> Distribution of files to ec sets is done by dht. In theory if you
>>>>>>>> create
>>>>>>>> many files each ec set will receive the same amount of files. However
>>>>>>>> when
>>>>>>>> the number of files is small enough, statistics can fail.
>>>>>>>>
>>>>>>>> Not sure what you are doing exactly, but a mapreduce procedure
>>>>>>>> generally
>>>>>>>> only creates a single output. In that case it makes sense that only
>>>>>>>> one
>>>>>>>> ec
>>>>>>>> set is used. If you want to use all ec sets for a single file, you
>>>>>>>> should
>>>>>>>> enable sharding (I haven't tested that) or split the result in
>>>>>>>> multiple
>>>>>>>> files.
>>>>>>>>
>>>>>>>> Xavi
>>>>>>>>
>>>>>>>>
>>>>>>>>>
>>>>>>>>> Thanks,
>>>>>>>>> Serkan
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> ---------- Forwarded message ----------
>>>>>>>>> From: Serkan Çoban <cobanserkan at gmail.com>
>>>>>>>>> Date: Mon, Apr 18, 2016 at 2:39 PM
>>>>>>>>> Subject: disperse volume file to subvolume mapping
>>>>>>>>> To: Gluster Users <gluster-users at gluster.org>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Hi, I have a problem where clients are using only 1/3 of nodes in
>>>>>>>>> disperse volume for writing.
>>>>>>>>> I am testing from 50 clients using 1 to 10 threads with file names
>>>>>>>>> part-0-xxxx.
>>>>>>>>> What I see is clients only use 20 nodes for writing. How is the file
>>>>>>>>> name to sub volume hashing is done? Is this related to file names
>>>>>>>>> are
>>>>>>>>> similar?
>>>>>>>>>
>>>>>>>>> My cluster is 3.7.10 with 60 nodes each has 26 disks. Disperse
>>>>>>>>> volume
>>>>>>>>> is 78 x (16+4). Only 26 out of 78 sub volumes used during writes..
>>>>>>>>>
>>>>>>>>
>>>>>>
>>>>
>>