SDSC HPSS Frequently Asked Questions

Introduction

This page is intended to answer some general questions that may arise in the SDSC user community concerning issues related to the use of the HPSS system. For example a major component of HPSS uses DCE and users are required to get DCE accounts. So this page will touch on some of these related issues. This is not intended to be a place to answer all questions or even contain all questions, but this will give you some hints of where to start.

You can also check here for a list of known HPSS problems.

• What is HPSS and what does HPSS stand for?
  
  HPSS stands for the HIGH PERFORMANCE STORAGE SYSTEM
  
  For an introduction as to what HPSS is, and an explanation of the product collaboration click here
  
  For a description of the HPSS project here at SDSC click here

• How do I get an account to use HPSS?
  
  If you don't already have an SDSC account, see SDSC New User Information
  

  If you already have an SDSC account then you need to be HPSS enabled. This will include having your HPSS home directory created as part of that process.
  Send account requests to allocations@npaci.edu
  

• How do I access the HPSS system and start storing data?
  
  You can use the HSI or pFTP interface to access HPSS.
  
  Go to NPACI User Guides under GETTING STARTED for documentation on using these user interfaces to HPSS.
  

• What does it mean when pftp puts out the message "Cannot load thread state"?
  
  This message generally means that a key HPSS server, such as the Bitfile Server, has crashed. When this happens, the pftp daemon cannot contact the HPSS servers, so login attempts and file transfers will fail. If this condition persists for more than a few minutes, please call the SDSC operators at (858) 534-5090 to report the problem.
  
  

• I am using passwordless pftp, but am being prompted for my password. Why?
  
  This can occur if:
  • You are not authorized to use pftp in passwordless mode
  • You are running on a host which is not allowed to use passwordless mode
  • One or more HPSS servers are down. This situation usually occurs in conjunction with the "Cannot load thread state" message (see above), and is usually an indication that HPSS servers have crashed.
  
  
  If you are getting messages on a host that is preventing you from accessing files on HPSS, please notify NPACI Consulting , or call the SDSC operators at (858) 534-5090.
  
  

• I am using the pftp interface to HPSS and I don't know the passwd, how do I find out what my passwd is?
  
  The version of pftpon the supercomputers is slightly different from the version on the SDSC worksations. On the supercomputers, pftpdoes not require you to enter a password. So you can use it interactively or as a single execute line. On the workstations, you must use interactive pftp. If you are still having problems Submit a Question to NPACI Consulting for more help.
  

• What is a "Class of Service", and how do I use it?
  
  A class of service (or "COS") is used to select a specific hierarchy of storage devices. HPSS supports multiple dynamic hierarchies, each of which contains one or more levels of devices with specific characteristics, such as high speed disk, memory, or tape. Each class of service is identified internally within HPSS by a small integer number, which is selected when the COS is defined by the site. Note that COS IDs are uniquely defined at each site.
  
  At SDSC, classes of service have been defined for small, medium, large, and extra-large files. Although it's possible to specify a particular class of service (see the pftp "setcos"), or hsi "cos" command), we have added code to the HPSS PFTP Daemon to automatically select a class of service based upon file size. If the file size cannot be determined before the transfer takes place, e.g., for ftp clients which do not support the ALLO command, or for files read from a Unix pipe, a temporary class of service is chosen for the duration of the transfer, and then the class of service is changed to the correct one based upon the final file size.
  
  
• What are the Classes of Service at SDSC?
  Note: these are subject to change and are listed here for information purposes only
  
  

  SDSC Classes of Service

  ID	File Type	File Sizes	Copies	Typical file types
  1	Small files	8,192 - 4 MB	1	Program sources, small html files, etc.
  2	Small files	0 - 4 MB	2	Industrial partner use
  4	Tiny files	0 - 8,192	1	Files that contain very little data, scripts, etc.
  10	Medium files	2 MB - 800 MB	1 & 2	Images, short animations, program binaries, etc. 2 copies for Critical files and Industrial partners
  20	Large files	800 MB - 42 GB	1 & 2	Program data files, small database table spaces. 2 copies for critical files and Industrial partner use.
  40	DB2	N/A	1	DOCT/MDAS project use (reserved COS)
  201	Very large files	42 GB - 500 GB	1	Very large backup files, etc. This is a direct-to-tape class of service using 1-way striping.
  202	Very large files (high-speed network access to HPSS)	6 - 500 GB	1	Very large files. This COS, which uses 2-way striped tape, should be used only from systems with a high enough data transfer rate to support streaming to 2 tape drives. Each drive can nominally support 7-8 MBS rates. This COS should be used with care.

  
  
  
• Why are HPSS "puts" faster than "gets"?
  
  This has to do with understanding how HPSS works. It is not unusual (actually it is quite normal) to see "puts" have megabyte transfer speeds from hsi or pftp command initiation to command end. Nor is it unusual to see slower timings from pftp command initiation to command end for a "get". For puts, the file transfers are cached to disk and transfer speeds can be quite high. For a "get" IF THE FILE IS STILL ON CACHE you can see the same transfer speeds as the "put". As files are created on cache they are soon after migrated to tape (so disk cache and tape copies exist). However, the cache is a finite resource broken up into Class Of Service which service different size files. As the cache(s) becomes full, files are purged from cache leaving only the tape copy. Then, if you perform an hsi or "pftp get" command for a file on tape and you time the transfer from command initiation to command end, you will most likely see slower transfer speeds. However, much of the time is accounted for in the MOUNTING OF THE TAPE AND POSITIONING THE TAPE to the file you request. The actual transfer speeds once the tape is in position is in the range you see on the puts (albeit a bit slower because of the tape speeds will be slower than disk speeds, but essentially the same as the drives are quite fast).
  The transfer time that PFTP reports is based upon the actual wall clock time between the file "open" and "close"; in the case of tape retrievals, it includes all of the following:
  
  - queue delay waiting for tape drive and/or tape cartridge(s) (Note that the large and extra-large Classes of Service use striped tapes).
  
  - any delay incurred while HPSS frees up disk cache space to allow the file to be copied (normally, this is not an issue, as the purge policies are set up to keep space available at all times - but it_can_happen).
  
  - tape mount time
  
  - positioning time
  
  - tape transfer time
  
  - disk transfer time
  
  Queue delay is extremely variable, particularly for tapes which are on the shelf and require operator intervention.
  
  It is also possible that HPSS will be ready to transfer the file and the client (the Cray for example) may be too busy to take it when HPSS is ready to transfer it. In this case HPSS will wait until the client is ready, which will add time to the total file transfer time.