Speaker
Robert Petkus
(Brookhaven National Laboratory)
Description
The RHIC/USATLAS Computing Facility at BNL has evaluated high-performance, low-cost
storage solutions in order to complement a substantial distributed file system
deployment of dCache (>400 TB) and xrootd (>130 TB). Currently, these file systems
are spread across disk-heavy computational nodes providing over 1.3 PB of aggregate
local storage. While this model has proven sufficient to date, the projected demand
for disk storage over the next five years is expected to cap 30 petabytes. At this
scale, more cluster nodes may be needed to meet storage requirements than necessary
for computation alone, introducing potential architectural inefficiencies. Candidate
solutions include extending the existing paradigm (more and larger disks per compute
node), augmentation with a cache of file servers, or consolidation of distributed
storage onto dedicated file servers. The product of our evaluation process will
strive to address these concerns and provide a clear path for the future.
To this end, a spectrum of storage solutions was subjected to both synthetic and
production benchmarking in an effort to discern the maximum I/O capacity of each
system. The goal was to identify the highest achieving and most versatile storage
systems, compare and contrast technology (e.g., RAID implementation, SATA vs. SAS
disks, Solaris/ZFS vs. Linux/ext3), and consider cost, ease of management and
scalability. Since data center real estate and power concerns are paramount,
additional consideration was given to ultra high-density storage systems. Test
methodology, results, and analysis are provided.
| Submitted on behalf of Collaboration (ex, BaBar, ATLAS) | RHIC/USATLAS |
|---|
Author
Robert Petkus
(Brookhaven National Laboratory)
Co-authors
Jason Smith
(Brookhaven National Laboratory)
Ofer Rind
(Brookhaven National Laboratory)
