While working with a large cloud provider leveraging HNAS (the core NAS platform at HDS)
for high-intensity cloud workloads, we found that the customer was very pleasantly surprised by
how well HNAS performed under heavy client loads. In fact, the Devops team for the customer had been
monitoring HNAS servers and became very concerned that these servers were running
at 100% of their possible load, and thus were very likely to go down, taking out a critical, customer-facing
service for their cloud:
Quoting from customer emails to our support team:
"Interestingly although [HNAS head] 5 is still pegged at 100% [the application] is not timing out and [HNAS's] drop duration
seems totally normal. Counter intuitive."
Although enterprise NAS customers generally focus on peak SPECsfs performance
and enterprise data center support, there are many customers (especially big ones)
that really need scalable file system performance under heavy load.
In particular, for customer-facing applications, it is critical that the system maintain sustained,
uniform performance, especially under heavy load. What does uniform performance mean? One
definition used by experienced system designers is:
Given a general level of performance under load x during period t, the subsystem should provide an
equivalent level of performance within a given small time delta, t + 1.
Uniform performance means that even as load increases, performance should either stay
the same, or increase if the system has additional headroom to add performance. Under no circumstances
should system performance drop unexpectedly or change in an unpredictable fashion.
Sustained, uniform performance brings many benefits:
- under heavy load and during critical service times, the system consistently delivers the sustained, uniform performance necessary to support the workload;
- system architects can confidently provision and plan for critical workloads given past system performance;
- it avoids inefficient provisioning due to the need for excessive headroom typically required for systems that cannot deliver uniform performance.
So why did HNAS surprise this customer? HNAS systems leverage FPGA technology to effectively partition NAS file processing into a set of independent pipeline steps executed on different pieces of hardware. In contrast, traditional NAS platforms run completely in software, creating unexpected performance bottlenecks as these pipeline steps execute in software against a shared memory system. HNAS partitioning separates the file and network processing into a set of independent steps, supporting uniform, sustained, predictable system behavior, especially under the heavy load conditions seen by this customer.
And the HNAS uniform, sustained performance story just keeps getting better. FLASH storage and new memory systems like Hybrid Memory Cubes are ravenous consumers and producers of memory bandwidth, so the efficient partitioning used by HNAS leverages these critical technologies even more to deliver the sustained, uniform performance so important to system designers. So if you want to build systems that only give you pleasant surprises while running the toughest, customer-facing workloads, give our HNAS platform a look.