HiPEAC voices Sweeping up best-paper awards and attracting widespread media coverage, Bianca Schroeder’s research is shaking up the field of storage devices and system reliability. HiPEAC caught up with the University of Toronto professor in advance of her keynote talk at HiPEAC 2022.
‘It is hard to imagine the capacity needs of the global datasphere being met entirely by SSDs’ Tell us about some of the main trends in
user demand for more capacity. So essentially the future of
storage over the last 10 years.
HDDs will depend on their ability to continue to offer capacity
In terms of hardware, a major trend has of
at a significantly lower $/GB than SSDs. In the next few years
course been the proliferation of flash-based
this could be through new methods (such as heat-assisted or
solid-state drives (SSDs). Also, more recently,
microwave-assisted magnetic recording, known as HAMR and
after decades of research, phase-change memory has finally made
MAMR) for increasing density.
its way from prototypes into products. While I think the jury is still out on what the main impacts of phase-change memory
What will probably change already in the near future is where
will be, SSDs have become ubiquitous everywhere ranging from
we will see HDDs deployed. Especially with the increased use
personal devices to datacentre and high-performance computing
of cloud storage, SSDs might pretty soon replace HDDs in
(HPC) deployments.
consumer devices. However, unless the $/GB for SSDs comes down significantly compared to HDDs it is hard to imagine that
In storage systems, disruptive innovation usually requires techni
the huge capacity needs of the global datasphere can be met
cal advances across the entire storage stack and the success story
entirely by SSDs. Remember that even tape is not dead yet! It is
of SSDs is no exception. At the device level, the key for success
still heavily used to meet the capacity needs of backups, archival
has been to achieve higher density with minimal impact on cost,
applications and cold storage.
by successfully squeezing more bits into a cell (from originally one bit in single-level cell (SLC) drives to most recently four
What techniques do you use for resource allocation and
bits in quad-level cell (QLC) drives) and stacking cells vertically
scheduling?
(3D NAND).
I have applied queueing theory to scheduling in web servers, database systems and high-performance computing systems in
Going up the storage stack, we have seen a new interface (NVMe), flash-optimized systems software (e.g. flash-aware filesystems with TRIM support ), and flash-optimized application software (e.g. key value stores based on flash-optimized data structures, such as log-structured merge trees), all of which have contributed to the success of SSDs. Seeing how SSDs have gone from use in only a few special performance critical systems 10-15 years ago to ubiquity today, it might be tempting to conclude that they will soon completely replace hard-disk drives (HDDs). And in fact, more SSDs are now being shipped annually than HDDs. However, the story changes when looking at capacity: the total capacity of shipped HDDs exceeds that of SSDs by a factor of five times and one of the few things that have been a constant in the history of storage systems seems to be the ever growing
order to improve response times and I am currently exploring the
