Industry executives and experts share their predictions for 2023. Read them in this 15th annual VMblog.com series exclusive.
What 2023 Might Look Like for Storage and the Cloud
By Boyan Ivanov, CEO of StorPool
New technologies and standards in storage and
server hardware
Compute Express Link (CXL), a CPU-to-device connection standard,
like PCI Express, is finally becoming available in standard x86 servers. In the
first generation of servers that support it, CXL will enable memory expansion
modules, so servers with several TB RAM will be economical. Previously RAM was
only available in DDR4 Registered DIMMs, which were limited in number and
capacity.
CXL persistent memory modules will supersede Optane NVDIMMs as the
most popular persistent memory technology because of wide compatibility and
easier integration.
And finally, CXL-connected "memory-semantic" SSDs will
make their first small steps on the market over time, becoming a mainstream
product like NVMe SSDs did in the previous decade.
Flexible Data Placement (FDP), a standard for SSDs for write
amplification reduction, will become more widely available in datacenter SSDs.
With it, the life of flash media is extended to support heavier workloads for
longer periods of time. FDP requires support from the storage software, so it
will be interesting to follow how it pans out.
QLC SSDs will grow their storage market share
for colder storage use cases
QLC SSDs continue to gain popularity in storage products that
serve "capacity-optimized" and "capacity-oriented" applications - i.e., data
lakes, analytics, backups, disaster recovery, and dev/test environments.
For primary workloads like heavily-loaded databases, virtual desktops,
and web servers, the 15%-20% lower cost per TB compared to TLC SSDs does not
justify the lower drive endurance, lower write performance, and higher read and
write latency. QLC SSDs simply do not provide sufficient value for money in the
context of primary storage systems yet. They will become attractive for primary
workloads as hardware vendors find ways to improve their endurance to at least
1 DWPD or introduce technologies like Flexible Data
Placement in next-generation QLC SSDs.
Legacy storage architectures continue facing
headwinds everywhere
The dual-controller shared-disk primary storage array designs and
the software-defined implementations of that architecture make sense for simple
deployments with predictable workload requirements. However, they are losing
ground both at the edge and in core data centers because today's environments
are neither simple nor predictable.
At the edge, small HCI solutions are pushing out standalone
storage systems thanks to their flexibility, ease of deployment and management,
and simple approach to high availability.
In core data centers, modern applications demand massive I/O
parallelization and low latency at scale. Fleets of standard servers are
preferred for medium- and large-scale infrastructure deployments. New clouds
use such building blocks to deliver structured and unstructured data storage
services. Customers need solutions that meet their performance needs, and
enable end-to-end automation and a wide range of hardware choices, while
deployments, upgrades, and refreshes happen on their timeframes. Continuous
workload monitoring, the need for workload rebalancing between data silos, and
performance degradation due to usable capacity consumption are all becoming
things of the past.
Storage services priced for performance and size
independently
Public clouds are starting to offer independent pricing of
capacity and performance. Previously the pricing of block storage services was
based on performance tiers. It is now moving to independent size (GiB
provisioned) and performance (IOPS, MB/s) pay-per-use pricing. Example services
are AWS EBS gp3 and Google Cloud HyperDisk. We expect that similar pricing of
storage services will be used by smaller service providers and managed storage
vendors.
Latency-optimized vs throughput-optimized CPUs
In public and private clouds, the split between latency-optimized
configurations (milliseconds query response time) and throughput-optimized
server configurations ($ per unit of work, lower power per core) is becoming
increasingly wider. For throughput-optimized workloads, ARM servers are getting
a bigger market share with Amazon's Graviton CPUs and Ampere Altra CPUs. X86
CPU vendors are also experiencing a split in their portfolio, with some
products targeting 2-3W per core for throughput-optimized workloads, while
other server CPUs use 10W+ per core for latency-optimized workloads.
In 2023 many public and private clouds will have two compute
offerings - one for batch processing and throughput-optimized workloads and a
second one for online transaction processing and latency-sensitive
workloads.
##
ABOUT THE AUTHOR
Boyan Ivanov is CEO of StorPool Storage,
a leading global storage software provider.