First published on MSDN on May 14, 2017
Guest blog by Wolfgang Gentzsch,
We are all very familiar with mass market economics: when things become available in abundance their prices tend to drop dramatically and they become affordable for everybody. And that's the secrete of the just announced
low-priority virtual machines
(VMs) to reduce the cost of Batch workloads. Low-priority VMs make new types of Batch workloads possible by providing a large amount of compute power that is also economical. Low-priority VMs are allocated from a surplus compute capacity and are available for up to an 80% discount, enabling certain types of workloads to run for a significantly reduced cost or allowing you to do much more for the same cost.
What is Azure Batch
It's a platform service for running large-scale parallel and high-performance computing (HPC) applications efficiently in the Azure cloud. Azure Batch schedules compute-intensive work to run on a managed collection of virtual machines, and can automatically scale compute resources to meet the needs of your jobs. Azure Batch is best suited for a large number of similar tasks which can be executed in parallel with no or minor interaction, for example, parameter studies in digital manufacturing to find the best suited material or geometry for a certain product; or rendering, analysis, and processing of thousands of images in digital content creation; or for analyzing genetic sequences in the life sciences
And as with any market dynamics, when things are becoming scarce again or they are becoming unavailable again, then prices are simply increasing, often drastically. In our case here the trade-off for using low-priority VMs is that those VMs may be preempted when no surplus capacity is available in Azure. For this reason, low-priority VMs are most suitable for certain types of workloads where the job completion time is flexible and the work is distributed across many VMs. Then, if a few VMs are needed for other - high-priority - work, the remaining low-priority virtual machines simply continue to work on the remaining 'low-priority' tasks, and the final result may just take a bit longer; usually no big deal. Another very nice features is auto-restarting a job when a VM gets reclaimed; this way you won't lose your intermediate results.
This new offering at up to 80% discount is especially well suited for start-ups and small and mid-size enterprises with big computing tasks who otherwise couldn't afford to throw a lot of expensive VMs onto the problem. This way, indeed, big computing is becoming affordable for everybody; and bundled with the fully automated self-service
UberCloud CAE packages
on the Azure Marketplace it is also becoming available and usable at your fingertips. Another jigsaw piece towards the democratization of HPC.