In high-end computing environments, remote file transfers of very large data sets to and from computational resources are commonplace as users are typically widely distributed across different organizations and must transfer in data to be processed and transfer out results for further analysis. Local transfers of this same data across file systems are also frequently performed by administrators to optimize resource utilization when new file systems come on-line or storage becomes imbalanced between existing file systems. In both cases, files must traverse many components on their journey from source to destination where there are numerous opportunities for performance optimization as well as failure. A number of tools exist for providing reliable and/or high performance file transfer capabilities, but most either do not support local transfers, require specific security models and/or transport applications, are difficult for individual users to deploy, and/or are not fully optimized for highest performance.
Shift is a framework for Self-Healing Independent File Transfer that provides high performance and resilience for local and remote transfers through a variety of techniques. These include end-to-end integrity via cryptographic hashes, throttling of transfers to prevent resource exhaustion, balancing transfers across resources based on load and availability, and parallelization of transfers across multiple source and destination hosts for increased redundancy and performance. In addition, Shift was specifically designed to accommodate the diverse heterogeneous environments of a widespread user base with minimal assumptions about operating environments. In particular, Shift is unique in its ability to provide advanced reliability and automatic single and multi-file parallelization to any stock command-line transfer application while being easily deployed by both individual users as well as entire organizations.
Shift includes the following features, among others:
- support for local, LAN, and WAN transfers
- drop-in replacement for both cp and scp (basic options only)
- tracking of individual file operations with on-demand status
- transfer stop and restart
- email notification of completion, errors, and warnings
- local and remote tar creation/extraction
- rsync-like synchronization based on modification times and checksums
- integrity verification of transfers with partial retransfer/resum to
rectify corruption
- detection of silent corruption between transfers of the same file
- throttling based on local and remote resource utilization
- automatic retrieval/release of files residing on DMF-managed file systems
- automatic striping of files transferred to Lustre file systems
- fully self-contained besides perl core and ssh
- automatic detection and selection of higher performance transports and
hash utilities when available including bbftp, mcp, msum, and rsync
- automatic many-to-many parallelization of single and multi-file
transfers with file system equivalence detection and rewriting
Shift is in active production at the NASA Advanced Supercomputing Facility (https://www.nas.nasa.gov/hecc/support/kb/entry/300) and has facilitated approximately 2.0M transfers over 3.0B files totalling 170 PB (as of Dec. 2020) since deployment in March 2012.
For full details of the Shift architecture, see https://pkolano.github.io/papers/resilience12.pdf and https://pkolano.github.io/papers/hust15.pdf. For installation details, see "INSTALL". For usage details, see "doc/shiftc.1" (in man page format, viewable with "nroff -man").
Questions, comments, fixes, and/or enhancements welcome.
--Paul Kolano paul.kolano@nasa.gov