- We’re open-sourcing DCPerf, a group of benchmarks that represents the various classes of workloads that run in information middle cloud deployments.
- We hope that DCperf can be utilized extra broadly by academia, the {hardware} business, and web firms to design and consider future merchandise.
- DCPerf is available now on GitHub.
Hyperscale and cloud datacenter deployments represent the biggest market share of server deployments on this planet as we speak. Workloads developed by large-scale web firms working of their datacenters have very completely different traits than these in excessive efficiency computing (HPC) or conventional enterprise market segments. Subsequently, server design issues, trade-offs and goals for datacenter use instances are additionally considerably completely different from different market segments and require a special set of benchmarks and analysis methodology. Present benchmarks fall wanting capturing these traits and therefore don’t present a dependable avenue to design and optimize fashionable server and datacenter designs.
Introducing DCPerf
Meta developed DCPerf, a group of benchmarks to symbolize the various classes of workloads that run in cloud deployments. Every benchmark inside DCPerf is designed by referencing a big utility inside Meta’s manufacturing server fleet.
We used a number of new strategies to make sure benchmark representativeness, starting from low-level {hardware} microarchitecture options to utility and library utilization profiles, to investigate manufacturing workloads and seize the vital traits of those workloads in DCPerf. Designing and optimizing {hardware} and software program on future server platforms utilizing these benchmarks willmore intently translate into improved effectivity of hyperscaler manufacturing deployments.
Over the previous few years, now we have repeatedly enhanced these benchmarks to make them appropriate with completely different instruction set architectures, together with x86 and ARM. We additionally validated that the benchmarks can be utilized to judge rising business traits, (e.g., chiplet-based architectures), and added assist for multi-tenancy in order that benchmarks can scale and make use of quickly growing core counts on fashionable server platforms.
Utilizing DCPerf to enhance Meta’s compute server designs
We’ve got been utilizing DCPerf internally, along with the SPEC CPU benchmark suite, for product analysis at Meta to make the precise configuration selections for our information middle deployments. DCPerf additionally helps us make early efficiency projections which might be used for capability planning, establish efficiency bugs in {hardware} and system software program, and collectively optimize the platform with our {hardware} business collaborators.
DCPerf gives a a lot richer set of utility software program variety and helps get higher protection indicators on platform efficiency versus current benchmarks corresponding to SPEC CPU. Attributable to these advantages, now we have additionally began utilizing DCPerf to help with our determination making course of on which platforms to deploy in our information facilities.
Enhancing state-of-the-art computing platforms with our {hardware} business collaborators utilizing DCPerf
Over the past two years now we have collaborated with main CPU distributors to additional validate DCPerf on pre silicon and/or early silicon setups to debug efficiency points and establish {hardware} and system software program optimizations on their roadmap merchandise. There have been a number of situations the place now we have been in a position to establish efficiency optimizations in areas corresponding to CPU core microarchitecture settings and SOC energy administration optimizations.
The graphic under reveals areas of HW/SW design the place now we have seen DCPerf being consultant of manufacturing utilization and being helpful for delivering related efficiency indicators and assist with optimizations in addition to areas of future work.
We’re grateful for our collaborators’ assist and contributions utilizing DCPerf to drive innovation in such an vital and complicated space and count on to proceed enhancing the benchmarks with new model releases over time to adapt to rising applied sciences.
Enabling improvements by means of open collaboration
As we speak, we’re open-sourcing DCPerf with the aim to create a collaborative and open supply reference benchmark that can be utilized to design, develop, debug, optimize, and enhance state-of-the-art in compute platform designs for hyperscale.
As an open supply benchmark suite, DCPerf has the potential to change into an business commonplace technique to seize vital workload traits of compute workloads that run in hyperscale datacenter deployments.
