[jira] [Created] (HADOOP-11501) Add NFSv3 Support for FileSystem
Gokul Soundararajan created HADOOP-11501: Summary: Add NFSv3 Support for FileSystem Key: HADOOP-11501 URL: https://issues.apache.org/jira/browse/HADOOP-11501 Project: Hadoop Common Issue Type: New Feature Components: fs, tools Reporter: Gokul Soundararajan Assignee: Gokul Soundararajan Priority: Minor I'm planning to add NFSv3 filesystem support for Hadoop. This will allow jobs to read/write data from a NFSv3 server. The idea is to support those use-cases where data is already located on NFSv3 and cannot be moved into HDFS (or other filesystems) due to various restrictions. I will make sure it passes all the HCFS tests as defined here: https://wiki.apache.org/hadoop/HCFS -- This message was sent by Atlassian JIRA (v6.3.4#6332)
Re: NFSv3 Filesystem Connector
Hi Steve, Thanks for the pointers. I looked at the FileSystemContractBaseTest and MainOperationsBaseTest and wrote the setUp() functions to setup NFS underneath. The extension of these test classes aren't very clean and I hope I can use and the rest of team's help in writing them correctly. In my unit tests, I setup a Mock NFS server (a really simple in-memory one) but it would be better if we can launch the HDFS+NFS one instead. The idea behind lazy flush is that NFS has commands for WRITE and COMMIT. The COMMIT will guarantee that any lazily written writes are actually written to stable (disk) storage. The optimization trick we do is to do the COMMIT when the stream is flushed or when it is closed. So the data will be stable after the stream is closed. The NFS implementation follows the NFSv3 standard. In fact, I have run it with Linux, Mac (how I write and test my code), and NetApp storage. I need to look into the Hadoop Kerberos stuff and it would be great if we can add that. (This is a work in progress and may be added in a later revision). Thanks, Gokul On Thu, Jan 15, 2015 at 3:39 AM, Steve Loughran wrote: > Gokul, > > What we expect from a filesystem is defined in (a) the HDFS code , (b) the > filesystem spec as derived from (a), and (c) contract tests derived from > (a) and (b) > > > http://hadoop.apache.org/docs/current/hadoop-project-dist/hadoop-common/filesystem/index.html > > There's a wiki page to go with this : https://wiki.apache.org/hadoop/HCFS > > Current policy on 3rd party FS support is pretty much: > > >1. Nothing in the hadoop codebase that isn't easy for anyone to test >2. Nothing in the hadoop codebase that isn't straightforward for anyone >to maintain >3. No FS support bundled in the Hadoop distribution that doesn't match >(1) or (2). Note that you are free to use Bigtop to build your own stack >with your own FS, which is what Redhat do with GFS, intel did with >LustreFS, etc. > > for a definition of "anyone" as "anyone who knows roughly what they are > doing and is familiar with the hadoop codebase" > > If your proposal had been for some 3rd party FS, the answer would be "keep > it in your own source tree". > > Except this isn't quite an FS, is it? It's an NFS client, which should be > able to talk to any NFSv3 server, including the standard Linux and OSX > ones, as well as the NFS support that you get with Hadoop. > > Is that right? That with this code I could run tests on my Linux box which > would verify the client works with NFS? > > > > > > On 15 January 2015 at 02:15, Gokul Soundararajan > wrote: > > > Hi Colin, > > > > Yeah, I should add the reasons to the README. We tried LocalFileSystem > when > > we started out but we think we can do tighter Hadoop integration if we > > write a connector. > > > Some examples include: > > 1. Limit over-prefetching of data - MapReduce splits the jobs into 128MB > > splits and standard NFS driver tends to over-prefetch from a file. We > limit > > the prefetching to the split size. > > > > Its more the MR asks the FS for the block size, and partitions are based on > that value. You'd want to return the prefetch-optimised value > > > > 2. Lazy write commits - For writes, we can relax the guarantees for > writes > > (and making it faster) and commit just before when the task ends. > > > > We haven't actually specified (yet) what output streams do. A key > requirement is that flush() persists, at least as far as the Java code is > aware of..there's the risk that the underlying OS can be lazy, and in a VM, > the virtual hardware can be even lazier. Note that the s3n, s3a and swift > break this rule utterly -this isn't considered satisfactory. > > > > 3. Provide for location awareness - Later, we can hook some NFS smarts > into > > getFileBlockLocations() (Have some ideas but not implemented them yet). > > > > > That could be good. SwiftFS does this with some extensions to OpenStack > swift that went in in sync with our code. > > Like I said, a new FS client for an external FS wouldn't get in to the > hadoop codebase. We care about HDFS, somewhat about file://, though > primarily for testing, and for integration with cloud storage. > > An NFS client though, one that can be tested and is general purpose to work > with implementations of NFSv3, and ideally integrates with the Hadoop > kerberos auth mechanism, that could be nice. > > > > > Hope this helps. > > > > Gokul > > > > > > > > On Wed, Jan 14, 2015 at 10:47 AM, Colin McCabe > > wrote: > >
Re: NFSv3 Filesystem Connector
Hi Colin, Yeah, I should add the reasons to the README. We tried LocalFileSystem when we started out but we think we can do tighter Hadoop integration if we write a connector. Some examples include: 1. Limit over-prefetching of data - MapReduce splits the jobs into 128MB splits and standard NFS driver tends to over-prefetch from a file. We limit the prefetching to the split size. 2. Lazy write commits - For writes, we can relax the guarantees for writes (and making it faster) and commit just before when the task ends. 3. Provide for location awareness - Later, we can hook some NFS smarts into getFileBlockLocations() (Have some ideas but not implemented them yet). Hope this helps. Gokul On Wed, Jan 14, 2015 at 10:47 AM, Colin McCabe wrote: > Why not just use LocalFileSystem with an NFS mount (or several)? I read > through the README but I didn't see that question answered anywhere. > > best, > Colin > > On Tue, Jan 13, 2015 at 1:35 PM, Gokul Soundararajan < > gokulsoun...@gmail.com > > wrote: > > > Hi, > > > > We (Jingxin Feng, Xing Lin, and I) have been working on providing a > > FileSystem implementation that allows Hadoop to utilize a NFSv3 storage > > server as a filesystem. It leverages code from hadoop-nfs project for all > > the request/response handling. We would like your help to add it as part > of > > hadoop tools (similar to the way hadoop-aws and hadoop-azure). > > > > In more detail, the Hadoop NFS Connector allows Apache Hadoop (2.2+) and > > Apache Spark (1.2+) to use a NFSv3 storage server as a storage endpoint. > > The NFS Connector can be run in two modes: (1) secondary filesystem - > where > > Hadoop/Spark runs using HDFS as its primary storage and can use NFS as a > > second storage endpoint, and (2) primary filesystem - where Hadoop/Spark > > runs entirely on a NFSv3 storage server. > > > > The code is written in a way such that existing applications do not have > to > > change. All one has to do is to copy the connector jar into the lib/ > > directory of Hadoop/Spark. Then, modify core-site.xml to provide the > > necessary details. > > > > The current version can be seen at: > > https://github.com/NetApp/NetApp-Hadoop-NFS-Connector > > > > It is my first time contributing to the Hadoop codebase. It would be > great > > if someone on the Hadoop team can guide us through this process. I'm > > willing to make the necessary changes to integrate the code. What are the > > next steps? Should I create a JIRA entry? > > > > Thanks, > > > > Gokul > > >
Re: NFSv3 Filesystem Connector
Hi Niels, Thanks for your comments. My goal in designing the NFS connector is *not* to replace HDFS. HDFS is ideally suited for Hadoop (otherwise why was it built?). The problem is that we have people who have PBs (10PB to 50PB) of data on NFS storage that they would like process using Hadoop. Such amount of data is both time-consuming and costly to move around; some have used Sqoop and Flume but it is still painful. To help these folks, we built the connector to enable Hadoop analytics on this data. As NFS is an open standard, we believe that it would benefit everyone who have this use case. Regarding the performance point, I hope you don't think a NFS storage server is a box with several disks and a single network connection. The latest generation storage servers are clustered storage systems that can have 17000+ drives , holding 100PBs, and support 64 10GbE ports on each cluster node. The NetApp spec sheet is here: http://www.netapp.com/us/products/storage-systems/fas8000/fas8000-tech-specs.aspx I hope this clarifies why we want to make this contribution. It is to unlock additional data that can be processed by Hadoop. Thanks, Gokul On Wed, Jan 14, 2015 at 3:14 AM, Niels Basjes wrote: > Hi, > > The main reason Hadoop scales so well is because all components try to > adhere to the idea around having Data Locality. > In general this means that you are running the processing/query software on > the system where the actual data is already present on the local disk. > > To me this NFS solution sounds like hooking the processing nodes to a > shared storage solution. > This may work for small clusters (say 5 nodes or so) but for large clusters > this shared storage will be the main bottle neck in the processing/query > speed. > > We currently have more than 20 nodes with 12 harddisks each resulting in > over 50GB/sec [1] of disk-to-queryengine speed and this means that our > setup already goes much faster than any network connection to any NFS > solution can provide. We can simply go to say 50 nodes and exceed the > 100GB/sec speed easy. > > So to me this sounds like hooking a scalable processing platform to a non > scalable storage system (mainly because the network to this storage doesn't > scale). > > So far I have only seen vendors of legacy storage solutions going in this > direction ... oh wait ... you are NetApp ... that explains it. > > I am no committer in any of the Hadoop tools but I vote against having such > a "core concept breaking" piece in the main codebase. New people may start > to think it is a good idea to do this. > > So I say you should simply make this plugin available to your customers, > just not as a core part of Hadoop. > > Niels Basjes > > [1] 50 GB/sec = approx 20*12*200MB/sec > This page shows max read speed in the 200MB/sec range: > > > http://www.tomshardware.com/charts/enterprise-hdd-charts/-02-Read-Throughput-Maximum-h2benchw-3.16,3372.html > > > On Tue, Jan 13, 2015 at 10:35 PM, Gokul Soundararajan < > gokulsoun...@gmail.com> wrote: > > > Hi, > > > > We (Jingxin Feng, Xing Lin, and I) have been working on providing a > > FileSystem implementation that allows Hadoop to utilize a NFSv3 storage > > server as a filesystem. It leverages code from hadoop-nfs project for all > > the request/response handling. We would like your help to add it as part > of > > hadoop tools (similar to the way hadoop-aws and hadoop-azure). > > > > In more detail, the Hadoop NFS Connector allows Apache Hadoop (2.2+) and > > Apache Spark (1.2+) to use a NFSv3 storage server as a storage endpoint. > > The NFS Connector can be run in two modes: (1) secondary filesystem - > where > > Hadoop/Spark runs using HDFS as its primary storage and can use NFS as a > > second storage endpoint, and (2) primary filesystem - where Hadoop/Spark > > runs entirely on a NFSv3 storage server. > > > > The code is written in a way such that existing applications do not have > to > > change. All one has to do is to copy the connector jar into the lib/ > > directory of Hadoop/Spark. Then, modify core-site.xml to provide the > > necessary details. > > > > The current version can be seen at: > > https://github.com/NetApp/NetApp-Hadoop-NFS-Connector > > > > It is my first time contributing to the Hadoop codebase. It would be > great > > if someone on the Hadoop team can guide us through this process. I'm > > willing to make the necessary changes to integrate the code. What are the > > next steps? Should I create a JIRA entry? > > > > Thanks, > > > > Gokul > > > > > > -- > Best regards / Met vriendelijke groeten, > > Niels Basjes >
NFSv3 Filesystem Connector
Hi, We (Jingxin Feng, Xing Lin, and I) have been working on providing a FileSystem implementation that allows Hadoop to utilize a NFSv3 storage server as a filesystem. It leverages code from hadoop-nfs project for all the request/response handling. We would like your help to add it as part of hadoop tools (similar to the way hadoop-aws and hadoop-azure). In more detail, the Hadoop NFS Connector allows Apache Hadoop (2.2+) and Apache Spark (1.2+) to use a NFSv3 storage server as a storage endpoint. The NFS Connector can be run in two modes: (1) secondary filesystem - where Hadoop/Spark runs using HDFS as its primary storage and can use NFS as a second storage endpoint, and (2) primary filesystem - where Hadoop/Spark runs entirely on a NFSv3 storage server. The code is written in a way such that existing applications do not have to change. All one has to do is to copy the connector jar into the lib/ directory of Hadoop/Spark. Then, modify core-site.xml to provide the necessary details. The current version can be seen at: https://github.com/NetApp/NetApp-Hadoop-NFS-Connector It is my first time contributing to the Hadoop codebase. It would be great if someone on the Hadoop team can guide us through this process. I'm willing to make the necessary changes to integrate the code. What are the next steps? Should I create a JIRA entry? Thanks, Gokul
