28 September 2016

Co-evolving data nodes

Bing! a mail comes in from our friends in the States saying look! here's someone in New Zealand who has set up iRODS node to GridFTP data to/from their site. It is a very detailed document yet it looks a lot like the DiRAC/GridPP data node document. They have solved many of the same problems we have solved, independently.

The basic idea is to have a node outside your institute/organisation which can be used to transfer data to/from your datastore/cluster. With a GridFTP endpoint, you could move data with FTS (as we do with DiRAC), people can use Globus (used by STFC's facilities, for example), or data can be moved to/from other e-infrastructures (such as EUDAT's B2STAGE) or EGI. Regardless of the underlying storage, there will be common topics like security, monitoring, performance, how to (or not to) firewall it, how to make it discoverable, etc. It could be the data node in a Science DMZ.

The suggestion is that we (= GridPP, DiRAC, and in fact anyone else who is willing and able) contribute to a detailed writeup which can be published as an OGF document (open access publishing for free!, and because GridFTP is an OGF protocol), either community practice or experiences - and then have a less detailed paper which could be submitted to a conference or published in a journal. 

22 September 2016

Upgrading and Expanding Lustre Storage (part4)

With a 1.5PB Lustre file system set up we now need to transfer our data from the old lustre system, conveniently also 1.5 PB in size, before we can put it into production.

Migration of Data:

It was found that is was not possible to mount both Lustre 1.8 and 2.8 on the same client, therefore migration of data had to be done via rsync between two clients mounting the different Lustre file systems. Setting up an rsync demon on the clients was found to be an order of magnitude quicker than using rsync over ssh for transferring data between the two clients. Hard links, ACLs and extended attributes are preserved by using the “-HAX” option when transferring data. Up to a dozen clients were utilised over the course of about six weeks to transfer 1.5PB of data between the old and new Lustre file systems. After the initial transfer the old and new systems were kept in sync with repeated rsync runs, remembering to use the “—delete” option to remove files that no longer existed on the live Lustre system. MD5 checksums were compared for a small random selection of files. The final transfer from the old to new Lustre took about a day, during which the file system was unavailable to external users. Then all clients were updated to the new Lustre version.  

Real World Experience:

With the new Lustre system put into production we then recommissioned the old system to create a 3PB Lustre file system. The grid cluster has about 4000 job slots in over 200 Lustre client compute nodes. The actual cluster is shown below. Note that the compute nodes fill the bottom 12U of every rack, where the air is cooler, and storage above them in the next 24U.
Real world performance over half a year, March to September 2016, is shown below. When all job slots are running grid “analysis” workloads, requiring access to data stored in Lustre, no slow down in job efficiency was observed. An average of 4.8 Gb/s is seen for reading data from Lustre and 1.6Gb/s for writing to Lustre (which is always done through StoRM).
However, in one case a local user simultaneously ran more than a 1500 jobs each accessing a very large number of small files, in this case BIOinformatics data, on Lustre and a slow down in performance was observed. Once the user was limited to no more than 500 jobs no further issues were seen. It is expected that accessing small files on the Lustre filesystem is not efficient [1] and should be avoided or limited where possible. A future enhancement to Lustre is planned that will enable small files to be stored on the MDS which should improve small file performance [1]. 

The Queen Mary Grid site major workload is for the ATLAS experiment which keeps detailed statistics of site usage. We are responsible for processing about 2.5% of all ATLAS data internationally and about 20% of data processed in the UK. Remote data transfer statics are shown below. Over the last 6 months ATLAS has transferred 2.39 PB of data into the cluster (top left plot), the weekly totals are shown in the top left plot, with a maxim for one week of 340TB (an average 4Gb/s). The bottom plot shows that 2.3PB has been sent to other grid sites around the world from Queen Mary.

Future Plans:
  • Double the Storage of the cluster to 6PB in 2018.
  • Consider an upgrade to Lustre 2.9 which will have bug LU1482 fixed and also provide additional functionality such as user and group ID mapping which would allow the storage to be used in different clusters. However Lustre 2.9 is SL/Centos7 only.
  • Upgrade OSS servers to SL/CentOS 7 from SL6. 
  • Examine the use of ZFS in place of hardware raid which might help mitigate very long raid rebuild times after replacement of a failed hard drive.

Over the past 4 Blogs we have shown a  successful major upgrade of Lustre. Including the specification, installation, configuration, migration of data, and operation of hardware and software.  

21 September 2016

Upgrading and Expanding Lustre Storage (part3)

In this post we will describe how we went about benchmarking and optimising our Lustre file system.

Performance Tuning:

A number of optimisation were made to improve the performance of the Lustre OSSs. To test these optimisation the IOzone [6] benchmarking program was used. IOzone is used to perform a variety of read and write tests. It is able to operate on a single server or on multiple clients at the same time. 

First it is useful to have an estimate of possible performance before undertaking benchmarking. The typical maximum sustained throughput of a single disk is quoted at approximately 200MB/s. For a 16 disk raid 6 array the maximum sustained throughput for a single server is expected to be 2.4GB/s (excluding the two parity disks). For a Lustre system made up of 20, Dell R730XDs, with 16 disks in each, this should scale to 56GB/s. However, each server is only connected with a 10Gb/s ethernet connection. Therefore the maximum sustained throughput obtainable is 25GB/s.

To test a single server IOzone was run with 12 threads (equal to the number of cpu cores) each transferring a file size of 24GB in chunks of 1024kB (iozone -e -+u -t 12 -r 1024k -s 24g -i0 -i1 -i 5 -i 8). As well as the standard sequential read and write tests, results were obtained for stride reads, and mixed workloads, which does reading and writing of a file with accesses being made to random locations within the file. The values were chosen to match the expected workload (i.e. the reading of large, GigaByte sized) to reduce cacheing effects and to match the 1024k buffer size used in Lustre network transfers. 

Using the BgFS Tips and Recommendations for Storage Server Tuning [7] as reference we applied different sets of optimisations to the storage server.  

Optimisation 1
echo deadline > /sys/block/sdb/queue/scheduler
echo 4096 > /sys/block/sdb/queue/nr_requests
echo 4096 > /sys/block/sdb/queue/read_ahead_kb
echo madvise > /sys/kernel/mm/redhat_transparent_hugepage/enabled
echo madvise > /sys/kernel/mm/redhat_transparent_hugepage/defrag

Optimisation 2 or (3), used in conjunction with optimisation 1, optimises the linux file system cacheing which is used by Lustre to help improve performance.   
echo 5(1) > /proc/sys/vm/dirty_background_ratio
echo 10(75) > /proc/sys/vm/dirty_ratio
echo 262144 > /proc/sys/vm/min_free_kbytes
echo 50 > /proc/sys/vm/vfs_cache_pressure

To reduce raid alignment complications the partition was made directly on to the storage device (e.g. /dev/sdb) taking into account the raid configuration (block size, stripe size and width). Lustre uses the EXT4 file system although it is possible to use ZFS instead.

The results of six different IOzone tests on a single server with different optimisations are shown in figure below (top). The results clearly show the benefits of applying optimisations to the OS to improve file system performance. As optimisation 1+3 show the highest throughput this has been applied to the Lustre file system. 

The single server tests were carried out for each of the 20 R730XD servers as a cross check of performance and as a check for hardware issues. All servers were found to produce similar performance.  A cross check for of the single server benchmark test, for optimisation 1 only, limiting the storage servers to only 2G RAM, to remove caching effects, was performed and results were found to be consistent with the results presented here.

A near complete 1.5 PB lustre file system with 20 Dell 730XD servers was created with up to 24 client nodes dedicated to the benchmark tests.  Lustre is set up such that individual files remain on a single OSS (i.e. there is no striping of files across OSSs). The well known Lustre clients tunes were included by default [1].
echo 256 > /proc/fs/lustre/osc/*/max_pages_per_rpc
echo 1024 > /proc/fs/lustre/osc/*/max_dirty_mb

For Lustre benchmarking using multiple clients IOzone is run with the “-+m filename” option to specify the client nodes (iozone -+m iozone_client_list_file -+h [IP of master IOzone node] -e -+u -t 10 -r 1024k -s 24g -i0 -i1 -i 5 -i 8). The figure above (bottom) shows the benchmark results for different number of clients. Each client has a 10Gb/s network connection so this sets the upper limit of the storage performance until we have more than 20 clients (black solid line). As the number of clients increase the performance first increases and then falls off for all but the initial write test. The maximum performance of the storage is seen with 18 clients. The anomalous reread result for 18 clients is reproducible and may to be due to client side cacheing effects. With 24 clients the mixed workload performance is below that for 8 clients. The reason for the fall off in performance for large number of active clients is probable due to contention for resource when seeking data on the file system, this would be less important for the initial writes tests. 

If we assume that a typical data analysis job uses 5MB/s and there is a maximum of 4000 job slots, then a throughput of the complete Lustre system of 20GB/s would be required for our cluster. The read performance measured of the benchmark Lustre system is of the order 15-20GB/s. The performance of the full Lustre file system, including 20 R730XDs and 70 R510s, is expected to be at least double that of the benchmarked system. If the real world workload is dominated by read type workflows, as is expected, then the full Lustre system should be able to provide the 20GB/s performance required.

NOTE: A number of network optimisations were deployed in production based on recommendations found on the faster data web site[8], for both data transfers within the cluster and for those done over the WAN by StoRM, these have not been benchmarked.

For the final part of this story we will discuss the real world Lustre system we have had in production for over 6 months.

[6] IOzone: 

[7] BeeGFS Tips and Recommendations for Storage Server Tuning: http://www.beegfs.com/wiki/StorageServerTuning

[8] ESnet Fasterdata Knowledge Base: 

20 September 2016

Upgrading and Expanding Lustre Storage (part2)

In the last post I Introduced Lustre and our history of use at the Queen Mary Grid site and then discussed the motivation and benefits of upgrading. In this post I will describe our hardware setup and the most important software configuration options. 

Hardware Choice and Setup:

In order to reduce costs the existing Lustre OSS/OSTs, made up of 70 Dell R510s, with 12 two or three TB hard disks in raid six, were reused, providing 1.5 PB of usable storage. An additional 20 Dell R730XDs, with 16 six TB disks in raid six was also purchased, providing 1.5 PB of usable storage, matching the size of the existing Lustre file system.  The Dell 730XDs have two Intel E5-2609 V3 processors and 64GB of RAM. Lustre is a “light” user of CPU resources on the OSS/OST and the E5-2609 processor is one of the cheapest CPUs available. Cost saving were also made by not utilising failover OSS/OSTs hardware which helped reduce costs by 40%!

However, the new MDS/MDT was set up in a resilient, automatic failover configuration utilising two Dell R630s connected to a MD3400 disk array. The Dell R630s have two Intel E5-2637 V3 processors and 256 GB RAM. The disk array has 12 600GB 15K SAS disks in RAID 10. Only one MDS/MDT is used in the cluster and the hardware has been specified as high as affordable. The automatic failover was configured using Corosync, Cman, Fence-agents and the Red Hat resource group manager (rgmanager) packages. Lustre itself has protection against the MDT being mounted by more than one MDS at a time.
All servers (storage, compute and service nodes) are connected to one of seven top of rack Dell S4810 network switches with a single 10Gb SFP+ Ethernet connection, which in turn are connected with multiple 40Gb QSFP+ connections to a distributed core switch made up of two Dell Z9000s in a Virtual Lan Trunk (VLT) configuration (figure 1).
As a result of design choices and several years of evolution in hardware the network connections from storage and compute servers are mixed in the top of rack switches. This has the advantage of balancing power and network IO [4] but at the expense of a more complicated hardware layout. 
Figure 1. Schematic of the Queen Mary Grid Cluster hardware layout.
Software Setup:

The Lustre software was installed on a standard SL6 OS configured server. A patch has been applied to Lustre due to a bug, LU1482 [1], causing incorrect interaction between Access Control Lists (ACLs) and the extended attribute permissions. This is required by StoRM as attributes are used to store checksums of every file which, after every gridftp transfer, are compared between source and destination. This bug is be fixed in the future 2.9 release of Lustre. 
The Lustre manual [1] describes in detail how to setup and configure a Lustre system.
The MDT is formatted and mounted on the MDS using the commands below. On the MDS add the “acl" option when mounting the MDT to ensure ACL and extended attributes support. For simplicity we install the Lustre ManaGement Server (MGS) on the MDS. The MGS will not be discussed further.

[root@mds05 ~] mkfs.lustre —fsname=lustre_1 --mgs --mdt --servicenode= --servicenode=  --index=0 /dev/mapper/mpathb
[root@mds05 ~]# cat /etc/fstab 
/dev/mapper/mpathb  /mnt/mdt lustre rw,noauto,acl,errors=remount-ro,user_xattr  0 0

On the OSS/OST You need to specify each of the MDSs when you configure a Lustre OSTs. Once each file system has been mounted it becomes visible to Lustre.

[root$sn100 ~]mkfs.lustre —fsname=lustre_1 --mgsnode=mds05@tcp0 --mgsnode=mds06@tcp0 --ost --index=0 /dev/sdb
[root@sn100 ~]# cat /etc/fstab 
/dev/sdb                /mnt/sdb                lustre  defaults        0 0

Lustre Clients need to know about both MDS/MGS nodes when mounting lustre in order to be able to fail over. Lustre is mounted as standard POSIX file system, of type lustre, on clients.

[root@cn200 ~]# cat /etc/fstab 
mds05@tcp0:mds06@tcp0:/lustre_1 /mnt/lustre_1    lustre  flock,user_xattr,_netdev 0 0

The file system mounted on a client appears as any normal file system, just bigger!

[~]$ df -h
Filesystem                       Size  Used Avail Use% Mounted on
mds05@tcp0:mds06@tcp0:/lustre_1  2.9P  2.1P  710T  75% /mnt/lustre_1

StoRM is used for remote data management for all Virtual Organisations (VOs) supported by the site and supports SRM, HTTP(S) and GridFTP. Most data is transferred via GridFTP and three GridFTP nodes were found to be needed to provide the capacity to fully utilise the 20GB WAN link. A standalone, readonly, installation of XRootD is deployed and is remotely usable by all site supported VOs using standard Grid authentication.   

19 September 2016

Upgrading and Expanding Lustre Storage (part1)

At the Queen Mary Grid site we are now running a Lustre file system of over 3PB using the most recent release (2.8). Lustre is an open source, POSIX compatible, clustered file system presented to the Grid using the StoRM Storage Resource Manager. Over the next few posts I would like to describe the recent major upgrade of the Lustre file system. I will: 
  • Introduce Lustre and our history of use at the Queen Mary Grid site and then discuss the motivation and benefits of upgrading; 
  • Describe our hardware setup and the most important software configuration options; 
  • Go into the testing and performance tuning of the file system as seen on the file server and the lustre client; 
  • Finally I will outline the data migration procedure and real world performance we have seen.


The Queen Mary WLCG tier two site has successfully operated a reliable, high performance, efficient, budget oriented storage storage solution, utilising Lustre[1] StoRM [2] and xrootd [3], since 2010 [4,5]. 
Lustre is a open-source(GPL), POSIX compliant, parallel file system used in over half of the worlds Top 500 supercomputers. Lustre is made up of three components: One or more Meta Data Servers (MDS) connected to one or more Meta Data Targets (MDT), which stores the namespace metadata such as filenames, directories and access permissions; One or more Object Storage Servers (OSS) connected to one or more Object Storage Targets (OST) which stores the actual files; and clients that access the data over the network using POSIX filesystem mounts. The network is typically either Ethernet or Infiniband.
StoRM (STOrage Resource Manager) is a scalable and file system independent storage manager service (SRM). It supports standard access and transfer protocols like HTTP(S), WebDAV and GridFTP. It it is designed to work on top of any POSIX filesystems with Access Control Lists(ACL) support such as Lustre.
Previously the Lustre storage file system at Queen Mary has undergone expansion from 300TB to 1.5PB, an upgrade of Lustre from version 1.6 to 1.8.X, a network upgrade from multiple 1Gb to 10Gb ethernet, and migration of the MDS and MDT to new hardware. This upgrade will involve new hardware, a complete reinstalation of every OS and Lustre software on every storage server (MDS/OSS) and a migration of data from the old Lustre to the new.

Motivation for Upgrade:

Last year it was decided that a major software and hardware upgrade was required. This was driven by several reasons: The need to upgrade the Operating system (OS) from Scientific Linux (SL)5 to a supported OS such as SL6 or CentOS7; Use a supported Lustre version compatible with SL6 or CentOS7; To take advantage of new software developments providing improved performance and reliability; Migrate to a new MDS/MDT with hardware in warranty; Double the storage capacity to over 3PB and allow for a doubling again before 2020.
Consideration was given to use of other open source file systems such as CEPH and GlusterFS. However, it was decided early on that local knowledge and experience with Lustre; its maturity, reliability and performance; clear long term development and support from Intel and others; and POSIX support made Lustre the obvious choice.
It is possible to buy a commercially supported solution but this was beyond the budget available. Therefore the specification, installation, configuration and operation of hardware and software had to 
be done by the site team.

next post: Hardware Choices and Software Setup

Some Useful References:

[1] Lustre:

[2] StoRM:

[3] XrootD:

[4] CHEP2012:
Scalable Petascale Storage for HEP using Lustre: Journal of Physics: C.J. Walker D.P. Traynor and A.J. Martin. Conference Series 396 (2012) 042063 

[5] CHEP2014:
Optimising network transfers to and from Queen Mary University of London, a large WLCG tier-2 grid site: C J Walker, D P Traynor, D T Rand, T S Froy and S L Lloyd. Journal of Physics: Conference Series 513 (2014) 062048 

07 September 2016

(Cloud) storage plugfest

A quick note to advertise the coming SNIA/CloudWatch cloud storage "plugfest" - physically in Santa Clara, CA, US, but remote attendance is possible: www.cloudplugfest.org. 19-22 Sep (19-21 is the SNIA storage developer event itself.)

These events are often very interesting, bringing together different components - and standards - and making them work together. Submit your work on any or all of CDMI, OCCI, CIMI, TOSCA, XACML/SAML/X.509.

20 April 2016

ZFS compression for LHC experiments data

An interesting feature of ZFS is that it supports transparent compression. Different to typical file compression, ZFS compression works on the record size/block size that it writes (which is variable in ZFS depending on the data and file size itself). Since it is important to have a fast compression/decompression algorithm to reduce the overhead compared to file access without compression, it can not be expected to get compression results similar to for example bzip in its highest compression level.  Also, the data files of the LHC experiments are ROOT files which already store data in a compressed format.

Therefore, I was not expecting any benefit of enabling compression on our servers, but since the newly implemented algorithm LZ4 has nearly no overhead even for non-compressible data, it shouldn't hurt to enable it.  Especially since our storage servers have Dual-CPUs with 12 cores each, running most of the time idle.

After enabling the default lz4 compression on 4 machines that were already migrated to ZFS and copying data on it, the first compression result looks like this:

tank-2TB  32.5T  8.73T  23.8T         -    15%    26%  1.00x  ONLINE  -
tank-8TB   116T  24.0T  92.0T         -    10%    20%  1.00x  ONLINE  -

NAME                    PROPERTY       VALUE  SOURCE
tank-2TB                compressratio  1.03x  -
tank-2TB/gridstorage01  compressratio  1.03x  -
tank-2TB/gridstorage02  compressratio  1.03x  -
tank-2TB/gridstorage03  compressratio  1.03x  -
tank-2TB/gridstorage04  compressratio  1.03x  -
tank-8TB                compressratio  1.03x  -
tank-8TB/gridstorage01  compressratio  1.03x  -
tank-8TB/gridstorage02  compressratio  1.03x  -
tank-8TB/gridstorage03  compressratio  1.03x  -
tank-8TB/gridstorage04  compressratio  1.03x  -
tank-8TB/gridstorage05  compressratio  1.04x  -
tank-8TB/gridstorage06  compressratio  1.03x  -
tank-8TB/gridstorage07  compressratio  1.03x  -
tank-8TB/gridstorage08  compressratio  1.03x  -
tank-8TB/gridstorage09  compressratio  1.03x  -
tank-8TB/gridstorage10  compressratio  1.03x  -
tank-8TB/gridstorage11  compressratio  1.03x  -

tank-2TB  32.5T  8.45T  24.0T         -    11%    26%  1.00x  ONLINE  -
tank-8TB   116T  24.1T  91.9T         -     7%    20%  1.00x  ONLINE  -

NAME                    PROPERTY       VALUE  SOURCE
tank-2TB                compressratio  1.03x  -
tank-2TB/gridstorage01  compressratio  1.03x  -
tank-2TB/gridstorage02  compressratio  1.03x  -
tank-2TB/gridstorage03  compressratio  1.03x  -
tank-2TB/gridstorage04  compressratio  1.04x  -
tank-8TB                compressratio  1.03x  -
tank-8TB/gridstorage01  compressratio  1.03x  -
tank-8TB/gridstorage02  compressratio  1.03x  -
tank-8TB/gridstorage03  compressratio  1.03x  -
tank-8TB/gridstorage04  compressratio  1.03x  -
tank-8TB/gridstorage05  compressratio  1.03x  -
tank-8TB/gridstorage06  compressratio  1.03x  -
tank-8TB/gridstorage07  compressratio  1.03x  -
tank-8TB/gridstorage08  compressratio  1.03x  -
tank-8TB/gridstorage09  compressratio  1.03x  -
tank-8TB/gridstorage10  compressratio  1.03x  -
tank-8TB/gridstorage11  compressratio  1.03x  -

tank-4TB   127T  9.05T   118T         -     3%     7%  1.00x  ONLINE  -

NAME                    PROPERTY       VALUE  SOURCE
tank-4TB                compressratio  1.03x  -
tank-4TB/gridstorage01  compressratio  1.03x  -
tank-4TB/gridstorage02  compressratio  1.03x  -
tank-4TB/gridstorage03  compressratio  1.04x  -
tank-4TB/gridstorage04  compressratio  1.02x  -
tank-4TB/gridstorage05  compressratio  1.03x  -
tank-4TB/gridstorage06  compressratio  1.03x  -
tank-4TB/gridstorage07  compressratio  1.03x  -
tank-4TB/gridstorage08  compressratio  1.03x  -
tank-4TB/gridstorage09  compressratio  1.03x  -
tank-4TB/gridstorage10  compressratio  1.04x  -
tank-4TB/gridstorage11  compressratio  1.03x  -
tank-4TB/gridstorage12  compressratio  1.02x  -
tank-4TB/gridstorage13  compressratio  1.03x  -
tank-4TB/gridstorage14  compressratio  1.03x  -

tank-2TB  63.5T  15.4T  48.1T         -    11%    24%  1.00x  ONLINE  -

NAME                    PROPERTY       VALUE  SOURCE
tank-2TB                compressratio  1.03x  -
tank-2TB/gridstorage01  compressratio  1.03x  -
tank-2TB/gridstorage02  compressratio  1.04x  -
tank-2TB/gridstorage03  compressratio  1.03x  -
tank-2TB/gridstorage04  compressratio  1.03x  -
tank-2TB/gridstorage05  compressratio  1.03x  -
tank-2TB/gridstorage06  compressratio  1.03x  -
tank-2TB/gridstorage07  compressratio  1.03x  -

Although there is not much data stored so far on each of the machines, this means we can still reduce the used disk space by some percent, 2-4% here depending on the file system and the data on it.
We have a bit more than 1PB disk storage in total on our site and the servers with 2TB disks provide about 50TB usable storage each. If we can get 4% compression for all the data, that would mean we could get nearly the space provided by one of the 2TB-disk servers additionally for free, without the cost of a new machine, power, extra disks,.... ! And that's just with the default compression while the compression level could also be tuned in ZFS...
This saving could be even bigger if we consider that in the future sites will also store more non-LHC data, like for LSST, which use a different and maybe uncompressed file format.
Another positive aspect of compression is that it reduces disk I/O since it needs to read less data blocks from disk.

It will be interesting to see how the compression rate will be after all our servers have been switch over  to ZFS.