Design considerations¶
Why zstash?¶
NERSC HPSS and other similar tape storage systems are optimized for file sizes of 100 to 500 GB. The retrieval of a large number of small files that are spread across dozens or hundreds of tapes is very inefficient because it requires multiple loads into tape drives and positioning the tape.
The recommended practice is thus to bundle smaller files into larger ones. zstash was created specifically for this purpose. While the existing ‘htar’ utility offers a similar functionality, it also suffers from limitations that make it not suitable for archiving E3SM model output (see below).
Features of zstash:
Files are bundled into standard tar files of a specified maximum size. The default maximum size is 256 GB, but it can be adjusted at creation using the
--maxsize
command line argument.Individual files are never split between two separate tar files. As a result, the size of tar files varies, but will generally be smaller than the specified maximum size. The only exception would be if individual files are larger than the target size.
Tar files are named sequentially using a six digit hexadecimal number (000000.tar, 000001.tar, …).
Tar files are first created locally on disk (under the zstash/ subdirectory) and then uploaded to HPSS. Similarly, when extracting files, tar files will be first downloaded from HPSS into the local cache (if they are not already present).
An sqlite3 index database is maintained (zstash/index.db) that contains information about every archived files, including:
full pathname relative to top-level archival directory.
size
modification time
md5 checksum
tar file in which the file is included
offset within this tar file (for faster extraction).
Database enables faster retrieval of individual files by locating in which tar file a specific file is stored, as well as its location (offset) within the tar file.
Checksums (md5) of input files are computed on-the-fly during archiving. For large files, this saves a considerable amount of time compared to separate checksumming and archiving steps.
File integrity is verified by computing checksums on-the-fly while extracting files.
Why not use ‘htar’?¶
Htar is an HPSS utility with many nice features. From the NERSC htar documentation page:
HTAR is a command line utility that creates and manipulates HPSS-resident tar-format archive files. It is ideal for storing groups of files in HPSS. Since the tar file is created directly in HPSS, it is generally faster and uses less local space than creating a local tar file then storing that into HPSS. Furthermore, HTAR creates an index file that (by default) is stored along with the archive in HPSS. This allows you to list the contents of an archive without retrieving it to local storage first. The index file is only created if the HTAR bundle is successfully stored in the archive.
This sounds great. In fact, some of these htar functionalities served as inspiration for zstash. However, htar also suffers from a number of limitations that were deemed to be show stoppers:
Member File Path Length
File path names within an HTAR aggregate of the form prefix/name are limited to 154 characters for the prefix and 99 characters for the file name. Link names cannot exceed 99 characters.
Member File Size
The maximum file size the NERSC archive will support is approximately 20 TB. However, we recommend you aim for HTAR aggregate sizes of several hundred GBs. Member files within an HTAR aggregate are limited to approximately 68GB.
Both the member file path length and file size were simply too restrictive for archiving output from E3SM simulations. Furthermore, htar lacks the ability to compute checksums of member files.
Why not just use ‘cput -R’?¶
This is the brute force approach: recursive conditional put. It will recursively transfer all your files to tape and only do so if the file isn’t already on tape.
While this approach may work, it is not recommended when dealing with very large number of (relatively small) files. Instead, NERSC recommends to group small files together :
HPSS is a tape system and responds differently than a typical file system. If you upload large numbers of small files they will be spread across dozens or hundreds of tapes, requiring multiple loads into tape drives and positioning the tape. Storing many small files in HPSS without bundling them together will result in extremely long retrieval times for these files and will slow down the HPSS system for all users.