FSISUS13: Can you complete workloads over more time while not violating your maximum SLA?
How do you avoid load spikes to reduce the provisioned capacity required for your workload?
Flattening the workload demand curve can help you to reduce the provisioned capacity for a workload and reduce its environmental impact. In other words, if you can afford to spread out the load over a longer period of time, rather than having a higher peak in a shorter span of time, then you lower the overall resource demand for the workload. By doing so, you lower the overall amount of provisioned capacity, and thus lower overall energy consumption to meet the workload's demand.
FSISUS13-BP01 Do not complete a customer transaction in the shortest time when not required by end users
Prescriptive guidance
If your workload does not have time-sensitive requirements, consider running them during times when public demand is lower. This distributes energy consumption to flatten the resource demand curve. Evaluate your workload requirements to assess if you are able to make this adjustment.
FSISUS13-BP02 Introduce jitter to your scheduled tasks
Prescriptive guidance
-
Assess if your scheduled tasks can be distributed to run at random times during an hour or throughout the day. This minimizes the highs of peak demand load and spreads it across the day instead. Avoid using the same start minute of scheduled tasks. Doing so creates high demand for resources at a specific time, which introduces stress on energy consumption. Staggering job start times avoids load spikes and creates time-flexible workloads.
-
Evaluate whether highly intensive computational workloads such as financial simulation can be spread over time and run fewer instances to maximize renewable energy availability. If a grid computing workload is using a third-party scheduler, prioritize workloads that need to provide calculations for regulators and trading desks that need information prior to markets opening, so workloads that are not urgent can be pushed off and worked on at a consistent rate to maximize renewable energy availability. Additionally, verify that a proper fault tolerance framework is implemented, as restarting a launch can increase launch time and energy consumption.
-
Use Amazon Simple Queue Service (Amazon SQS)
achieve your goal. -
Balance generative AI model response time requirements with energy efficiency.
-
Implement cost-aware prompting strategies that may take slightly longer but use fewer resources.
-
Use distributed generative AI inference when time permits to optimize resource utilization.
