A location to store the data: the physical storage.
A way to implement reliability for the data: RAID.
A way to provide access to the data: a file system.
On June 20, 2007, the first successful demonstration of trans-Atlantic streaming over
photonic IP networks of 4K digital motion pictures and 5.1 surround sound was achieved
by the international research consortium, CineGrid.
3.3 RAID
• RAID-0: known as striping. A minimum of two disks is configured in one array [A]. Data
is read and written from and to all disks. This increases performance a lot. Disadvantage is
that if just one disk fails, all data on all participating disks is lost.
• RAID-1: known as mirroring. Data is written to two disks: one disk is the backup. Reliability
is increased at the cost of half of the capacity.
• RAID-5: parities are used to increase reliability. A minimum of three disks is required. At
each participating disk a parity block is used to recover data when a disk fails. For critical
applications often RAID-5 is used.
• RAID-6: similar to RAID-5, but at each disk are two parity blocks stored. This means
two disks can fail and still all data can be recovered. This increases reliability even more.
Disadvantage is that RAID-6 requires more processing power, costs and space.
———————–
Design Criteria
To design a storage architecture, a list of criteria must be made. The following criteria are important:
bandwidth, IOs per second (IOPS), budget, reliability, space, scalability, administration,
response time, complexity, maturity and flexibility.
———————–
Bandwidth
Bandwidth is an important criterion in this project. To determine the bandwidth that is needed,
we should look at the material that needs to be transferred: that would be 4K video. To calculate
the bandwidth, we used the following parameters:
• HZP (Horizontal pixels per frame) = 4096
• VTP (Vertical pixels per frame) = 2160
• CBP (Colour bits per pixel) = 3 bytes (Red, Green, Blue) = 3 x 8 bits = 24 bits
• FPS (Frames per second) = 30
Now we can calculate the bandwidth:
Bandwidth(Gb/s) = (HZP x VTP xCBP x FPS)/(109) = (4096x2160x24x30)/(109) = 6.4 Gb/s
Sound is not taken into account in this calculation!
About compression (see related work about DXT compression [6]): the bandwidth of 6.4 Gb/s is
a minimum requirement. There must be some headroom and the choice for multiple streams in
the future should be an option. This project is about the storage architecture: the data has to be
stored in real-time. Compression is a possibility in a later stage and is not part of this project.
Note: b means bit and B means Byte (8 bits).
需要带宽6.4Gb/8=0.8GB
6.4Gb/s 6370099 Kb/s = 777600KB/s
—————–
Space
To determine the needed space, we need the calculated bandwidth and the duration of the video
material. To calculate the space of one hour of 4K video material, we use the following formula:
Space(TB/hour) = (Bandwidth(Gbps) 109 Time(s))/(8 10244) =
(6.4 x 109 x 60 x 60)/(8 x 10244) = 2.6 TB/hour
This is the space that is needed for the proprietary umf format, which was talked about in 4.1. If
a two-hour movie has to be stored, then 5.2 TB would be needed. Normally in each year four to
six performances are stored. This means a total of 6 * 5.2 TB = 31.2 TB/year.
Note: b means bit and B means Byte (8 bits).
_________________________
File systems
• Ceph: a distributed, parallel, fault tolerant file system by the Storage Systems Research
Center at the University of California, Santa Cruz. Fault tolerance is realized by data replication.
When new nodes join the cluster, all data is automatically distributed to the new
nodes. Ceph is installed at kernel-level. Although Ceph looks promising, it is still a prototype
at this moment. Ceph is open source and falls under the GPL license. Website:
http://ceph.sourceforge.net/.
• Gfarm Grid File System: a distributed, parallel, fault tolerant file system by ApGrid. Gfarm
is installed at user-level with the help of FUSE. Gfarm has a stable version, but there are
known bugs and the program still is in development. Gfarm is open source and falls under
the X11 license. Website: http://datafarm.apgrid.org/.
• GFS: Global File System by Red Hat. GFS is a shared disk file system available with Red
Hat cluster suite. Only a SAN and only Linux Red Hat is supported. GFS is open source
and falls under the GPL license. Website: http://www.redhat.com/gfs/.
• GlusterFS: a distributed, parallel, fault tolerant file system by Z RESEARCH. GlusterFS is
installed at user-level on top of an existing file system. This makes it very flexible in use.
No single point of failure (SPOF) exists and all metadata is handled by the underlying file
system. Performance is good. Installation is easy. GlusterFS is open source and falls under
the GPL license. Website: http://www.gluster.org/.
• Google File System: a distributed, parallel, fault tolerant file system by Google. This file
system is only used internally at Google. Website: http://labs.google.com/papers/gfs.
html.
• GPFS: General Parallel File System by IBM. GPFS is a proprietary shared disk file system.
Website: http://www.ibm.com/systems/clusters/software/gpfs.html.
• Hadoop Distributed File System: a distributed, parallel, fault tolerant file system by Apache.
Hadoop DFS runs on JAVA. This software is open source and the license is free. Website:
http://lucene.apache.org/hadoop/.
• Lustre: a distributed, parallel file system by Cluster File Systems. There is no replication of
data: reliability must be handled in external hardware. Performance is good. This product
is very mature and used on several supercomputers. Documentation and support are good.
Lustre is installed at kernel-level. The software still is open source, but it may be possible
proprietary add-ons will be developed in the future. Website: http://www.lustre.org/.
• NFS: Network File System. A distributed file system that comes with the Linux kernel.
Performance is less than with distributed parallel file systems. NFS is open source and falls
under the GPL license. Website: http://nfs.sourceforge.net/.
• OpenAFS: Open Andrew File System by IBM. This is a distributed file system with read-only
replication. Limits with OpenAFS are not as high as with the other file systems: 8 GB per
volume is supported and the maximum file size is 2 GB. IPv6 is not supported. OpenAFS is
open source and falls under the IBM Public License. Website: http://www.openafs.org/.
• OCFS: Oracle Cluster File System. OCFS is a shared file system by Oracle. OCFS is
open source and falls under the GPL license. Website: http://oss.oracle.com/projects/
ocfs2/.
• PeerFS: a distributed, parallel, fault tolerant file system by Radiant Data Corp. In the area
of reliability the software only supports mirroring. PeerFS is proprietary software. Website:
http://www.radiantdata.com/English/Products/PeerFS.html.
• PVFS: Parallel Virtual File System. A distributed, parallel file system by several contributors.
PVFS is installed
at user-level. The software is open source and falls under the GPL
license. Website: http://www.pvfs.org/.
• TerraGrid: a distributed, parallel, fault tolerant file system by Terrascale. This software is
proprietary. Development is not active anymore. Website: http://www.terrascale.com/.
Post comment