High Performance Storage System

HPSS Logo
Incremental Scalability
Based on storage needs and deployment schedules, HPSS scales incrementally by adding computer, network and storage resources. A single HPSS namespace can scale from petabytes of data to exabytes of data, from millions of files to billions of files, and from a few file-creates per second to thousands of file-creates per second.
About HPSS   :    HPSS RAIT Technology
What is HPSS RAIT?

HPSS RAIT (Redundant Array of Independent Tapes) is an HPSS software technology for striping data on tape with the added benefit of parity for redundancy. HPSS RAIT with single parity is similar to RAID-5 and HPSS RAIT with two parities is similar to RAID-6, except HPSS RAIT is used with tape. HPSS with RAIT still streams the tape drives at hardware speed, and stripe widths of up to 16 (data+parity) are available.

HPSS RAIT is transparent to the client using any of the existing HPSS client interfaces: FUSE, SwiftOnHPSS, FTP, Parallel FTP, HSI, HTAR, and other third-party HPSS client applications.

Why use HPSS RAIT?

HPSS RAIT offers two main advantages. First, RAIT allows data to flow to and from tape in a parallel manner. Parallel tape I/O allows single file transfers to exceed the native transfer rate of a single tape drive. At Oakridge National Laboratory (ORNL), large file I/O exceeds 1 GB/s with their 4+P HPSS RAIT stripe deployed on 250 MB/s tape technology. Thus, a 1 TB file can be written to tape in less than 17 minutes. Without parallel I/O, a 1 TB file might take over an hour to transfer using a single 250 MB/s tape drive.

Additionally, when compared to dual-copy tape, HPSS RAIT lowers redundant tape costs. At the lowest level, a 2+P HPSS RAIT will deliver up to 2x improvement in tape bandwidth and typically cuts the cost of redundant tape by 45%. Redundant tape costs are cut by 45%, rather than 50%, because typically 90% of the files by count only take 10% of the space in an HPSS repository, while 10% of the files by count take up 90% of the space. HPSS RAIT is intended for large files, so for greater recall efficiency, dual-copy tape is used to protect smaller files.

There are two basic methods of redundancy: making multiple copies (mirroring), and redundancy through error correcting codes (such as Reed-Solomon). RAIT uses Reed-Solomon error correcting codes which is less expensive than making multiple copies due to the lower tape, drive, space, and power requirements.

The Value of RAIT over Mirroring

There is a lower cost of fault tolerance with RAIT: less tape drives and cartridges are required. There are also faster mount times when compared with writing to a mirrored volume. For example, in the case of an 8 tape stripe and adding single failure redundancy: mirroring would require 8 additional tapes, resulting in a total of 16 tapes. RAIT, however, would only need 1 extra tape, resulting in a total of 9 tapes to accomplish the redundancy.

With RAIT, there is a high tolerance for tape loss within the stripe. There is transparent recovery in the event of a tape mount or media error. HPSS will continue to write to a stripe after the loss of one or more tapes, and will continue to read from a failed tape by reconstructing the missing data from the remaining parity.

What's Under the Hood of HPSS RAIT?

At the heart of HPSS RAIT is the HPSS RAIT engine. Client data will automatically flow through the HPSS RAIT engine for all HPSS RAIT transfers to and from tape.

Like HPSS' RAID counterparts, the HPSS RAIT parity rotates across all tapes in the RAIT stripe. Rotating the parity evens out the compressibility of the data on tape by ensuring an equal amount of data and parity are written to each tape cartridge.

The figure below illustrates a single HPSS RAIT engine accept four data blocks (A, B, C, and D) at a time from the client, generate the two parity blocks (P and Q), and send the six data blocks to tape in a parallel manner. Notice that the figure illustrates how the "File Data Block" is processed in four iterations, as follows:

1. A1, B1, C1, and D1 + P1 and Q1

2. A2, B2, C2, and D2 + P2 and Q2

3. A3, B3, C3, and D3 + P3 and Q3

4. A4, B4, C4, and D4 + P4 and Q4

The following illustration shows how data movement occurs:

HPSS RAIT Write Operations

HPSS RAIT Configuration Options and Features

  • HPSS RAIT supports data stripe width of up to 15 and parity stripe width of up to 7.
  • You can control whether to mount the entire RAIT volume, or only a minimal set of tapes for reading from a RAIT volume. For an 8 data/3 parity volume, this would cause 9 tapes to be mounted rather than 11 if the data is being verified against parity.
  • Read verification allows users to verify that the data being read still matches the parity, with stripe-level data integrity.

Proven in the Field

NCSA:
  • NCSA began using HPSS RAIT during a cycle of pre-production testing where they tested ingesting 5 billion files in a single namespace with constant ingest and retrieval.
  • During this testing, they ingested 426TB and retrieved 499 TB of data in 24 hours.
  • NCSA has been using HPSS RAIT in production since 2013.
  • NCSA has been able to ensure data integrity at one-fifth the cost of data mirroring.
ORNL:
  • ORNL began using HPSS RAIT in January 2015, and are seeing faster tape mounts when compared to writing a mirror.
  • ORNL was able to cut redundant tape cost-estimates by 75% with 4+Parity HPSS RAIT (tape stripe with rotating parity) and enjoy large file tape transfers beyond 1 GB/s.

< Home

Come meet with us!
HPSS @ MSST 2018
The 34th International Conference on Massive Storage Systems and Technology will be in Santa Clara, California from May 14th through May 18th, 2018 - Learn More. Please contact us if you would like to meet with the IBM business and technical leaders of HPSS at Santa Clara University.

HPSS @ ISC18
The 2018 international conference for high performance computing, networking, and storage will be in Frankfurt, Germany from June 24th through 28th, 2018 - Learn More. Come visit the HPSS folks at the IBM booth and contact us if you would like to meet with the IBM business and technical leaders of HPSS in Frankfurt.

2018 HUF
The 2018 HPSS User Forum will be hosted by the UK Met Office in Exeter, United Kingdom from October 15th through October 18th, 2018. Please contact us if you would like to meet with the HPSS Collaboration leaders, HPSS Collaboration developers, architects, system engineers, support staff, project managers, and many of the HPSS customers.

The 2018 HUF information and registration web site is coming soon.

HPSS @ SC18
The 2018 international conference for high performance computing, networking, storage and analysis will be in Dallas, Texas from November 11th through 15th, 2018 - Learn More. Come visit the HPSS folks at the IBM booth and contact us if you would like to meet with the IBM business and technical leaders of HPSS in Dallas.
What's New?
HPSS Meltdown and Spectre Vulnerability Statement - Learn More.

HPSS Vendor Partnership Grows - HPSS begins Quantum Scalar i6000 tape library testing in 2018. Other HPSS tape vendor partners include IBM, Oracle, and Spectra Logic.

Swift On HPSS - Leverage OpenStack Swift to provide an object interface to data in HPSS. Directories of files and containers of objects can be accessed and shared across ALL interfaces with this OpenStack Swift Object Server implementation - Contact Us for more information, or Download Now.

Capacity Leader - ECMWF (European Center for Medium-Range Weather Forecasts) has a single HPSS namespace with 360 PB spanning 328 million files.

File-Count Leader - LLNL (Lawrence Livermore National Laboratory) has a single HPSS namespace with 41 PB spanning 1.084 billion files.

RAIT - Oak Ridge National Laboratory cut redundant tape cost-estimates by 75% with 4+P HPSS RAIT (tape stripe with rotating parity) and enjoy large file tape transfers beyond 1 GB/s.
Home    |    About HPSS    |    Services    |    Contact us
Copyright 2018, HPSS Collaboration. All Rights Reserved.