Performance Evaluation for Tape Storage Data Recall with T10KD Drive

Tape library storage is the best solution for archiving big data due to high capacity, low cost and long retention. However, as the nature of sequential media, its main drawback versus disk storage is the slow performance for data recall in random access order. With emerging of the modern technology from the enterprise tape drives, Recommended Access Order (RAO) should improve the data recall performance considerably.The Scientific Data and Computing Center (SDCC) in Brookhaven National Lab (BNL) is operating a tape library storage system which contains more than 100 PB data. It is managed by High Performance Storage System (HPSS) software. Starting from HPSS 7.5.1 release, a new feature named Tape Ordered Recall (TOR) has offered a great advantage for large amount of data recall. It is designed to reduce the file access time from tape media. In our test with Oracle T10K enterprise drive, TOR's tape recall rates for small files (1GB) is almost 4 times of the unordered recall. TOR supports both RAO and non-RAO drives. It will use the RAO feature if available. Otherwise, it makes a linear offset ordering to recall the data. Additionally, Quaid is a smart recall tool introduced in HPSS7.5.1 as well. It uses new batch stage API to stage files in the background, improves the tape data recall performance significantly. Efficient Retrieval and Access to Data Archived on Tape, aka ERADAT, is a tape storage data recall tool developed mainly in SDCC. It sorts the staging requests based on file's logical position order before submitting jobs to HPSS. Over a decade of utilization and improvement, ERADAT has been proven as a great performer serving BNL's worldwide scientific community. In this paper we will present the test results for each individual tool under certain test setting and configuration, analyze and evaluate their performance advantage and disadvantage, find the best solution for BNL's tape storage production system for different data characteristics.