Search for: All records

Generative AI, exemplified in ChatGPT, Dall-E 2, and Stable Diffusion, are exciting new applications consuming growing quantities of computing. We study the compute, energy, and carbon impacts of generative AI inference. Using ChatGPT as an exemplar, we create a workload model and compare request direction approaches (Local, Balance, CarbonMin), assessing their power use and carbon impacts. Our workload model shows that for ChatGPT-like services, in- ference dominates emissions, in one year producing 25x the carbon-emissions of training GPT-3. The workload model characterizes user experience, and experiments show that carbon emissions-aware algorithms (CarbonMin) can both maintain user experience and reduce carbon emissions dramatically (35%). We also consider a future scenario (2035 workload and power grids), and show that CarbonMin can reduce emissions by 56%. In both cases, the key is intelligent direction of requests to locations with low-carbon power. Combined with hardware technology advances, CarbonMin can keep emissions increase to only 20% compared to 2022 levels for 55x greater workload. Finally we consider datacenter headroom to increase effectiveness of shifting. With headroom, CarbonMin reduces 2035 emissions by 71%. 

Cloud providers are adapting datacenter (DC) capacity to reduce carbon emissions. With hyperscale datacenters exceeding 100 MW individually, and in some grids exceeding 15% of power load, DC adaptation is large enough to harm power grid dynamics, increasing carbon emissions, power prices, or reduce grid reliability. To avoid harm, we explore coordination of DC capacity change varying scope in space and time. In space, coordination scope spans a single datacenter, a group of datacenters, and datacenters with the grid. In time, scope ranges from online to day-ahead. We also consider what DC and grid information is used (e.g. real-time and day-ahead average carbon, power price, and compute backlog). For example, in our proposed PlanShare scheme, each datacenter uses day-ahead information to create a capacity plan and shares it, allowing global grid optimization (over all loads, over entire day). We evaluate DC carbon emissions reduction. Results show that local coordination scope fails to reduce carbon emissions significantly (3.2%–5.4% reduction). Expanding coordination scope to a set of datacenters improves slightly (4.9%–7.3%). PlanShare, with grid-wide coordination and full-day capacity planning, performs the best. PlanShare reduces DC emissions by 11.6%–12.6%, 1.56x–1.26x better than the best local, online approach’s results. PlanShare also achieves lower cost. We expect these advantages to increase as renewable generation in power grids increases. Further, a known full-day DC capacity plan provides a stable target for DC resource management. 

The historical motivation for serverless comes from internet-of-things, smartphone client server, and the objective of simplifying programming (no provisioning) and scale-down (pay-for-use). These applications are generally low-performance best-effort. However, the serverless model enables flexible software architectures suitable for a wide range of applications that demand high-performance and guaranteed performance. We have studied three such applications - scientific data streaming, virtual/augmented reality, and document annotation. We describe how each can be cast in a serverless software architecture and how the application performance requirements translate into high performance requirements (invocation rate, low and predictable latency) for the underlying serverless system implementation. These applications can require invocations rates as high as millions per second (40 MHz) and latency deadlines below a microsecond (300 ns), and furthermore require performance predictability. All of these capabilities are far in excess of today's commercial serverless offerings and represent interesting research challenges. 

The rapid growth of datacenter (DC) loads can be leveraged to help meet renewable portfolio standard (RPS, renewable fraction)targets in power grids. The ability to manipulate DC loads over time(shifting) provides a mechanism to deal with temporal mismatch between non-dispatchable renewable generation (e.g. wind and solar) and overall grid loads, and this flexibility ultimately facilitates the absorption of renewables and grid decarbonization. To this end, we study DC-grid coupling models, exploring their impact on grid dispatch, renewable absorption, power prices, and carbon emissions.With a detailed model of grid dispatch, generation, topology, and loads, we consider three coupling approaches: fixed, datacenter-local optimization (online dynamic programming), and grid-wide optimization (optimal power flow). Results show that understanding the effects of dynamic DC load management requires studies that model the dynamics of both load and power grid. Dynamic DC-grid coupling can produce large improvements: (1) reduce grid dispatch cost (-3%), (2) increase grid renewable fraction (+1.58%), and (3) reduce DC power cost (-16.9%).It also has negative effects: (1) increase cost for both DCs and non-DC customers, (2) differentially increase prices for non-DC customers, and (3) create large power-level changes that may harm DC productivity. 

Datacenter scheduling research often assumes resources as a constant quantity, but increasingly external factors shape capacity dynamically, and beyond the control of an operator. Based on emerging examples, we define a new, open research challenge: the variable capacity resource scheduling problem. The objective here is effective resource utilization despite sudden, perhaps large, changes in the available resources. We define the problem, key dimensions of resource capacity variation, and give specific examples that arise from the natural world (carboncontent, power price, datacenter cooling, and more). Key dimensions of the resource capacity variation include dynamic range, frequency, and structure. With these dimensions, an empirical trace can be characterized, abstracting it from the many possible important real-world generators of variation. Resource capacity variation can arise from many causes including weather, market prices, renewable energy, carbon emission targets, and internal dynamic power management constraints. We give examples of three different sources of variable capacity. Finally, we show variable resource capacity presents new scheduling challenges. We show how variation can cause significant performance degradation in existing schedulers, with up to 60% goodput reduction. Further, initial results also show intelligent scheduling techniques can be helpful. These insights show the promise and opportunity for future scheduling studies on resource volatility. 

Cameras are deployed at scale with the purpose of searching and tracking objects of interest (e.g., a suspected person) through the camera network on live videos. Such cross-camera analytics is data and compute intensive, whose costs grow with the number of cameras and time. We present Spatula, a cost-efficient system that enables scaling cross-camera analytics on edge compute boxes to large camera networks by leveraging the spatial and temporal cross-camera correlations. While such correlations have been used in computer vision community, Spatula uses them to drastically reduce the communication and computation costs by pruning search space of a query identity (e.g., ignoring frames not correlated with the query identity’s current position). Spatula provides the first system substrate on which cross-camera analytics applications can be built to efficiently harness the cross-camera correlations that are abundant in large camera deployments. Spatula reduces compute load by $8.3\times$ on an 8-camera dataset, and by $23\times-86\times$ on two datasets with hundreds of cameras (simulated from real vehicle/pedestrian traces). We have also implemented Spatula on a testbed of 5 AWS DeepLens cameras.