Search for: All records

Recent innovation in large language models (LLMs), and their myriad use cases have rapidly driven up the compute demand for datacenter GPUs. Several cloud providers and other enterprises plan to substantially grow their datacenter capacity to support these new workloads. A key bottleneck resource in datacenters is power, which LLMs are quickly saturating due to their rapidly increasing model sizes.We extensively characterize the power consumption patterns of a variety of LLMs and their configurations. We identify the differences between the training and inference power consumption patterns. Based on our analysis, we claim that the average and peak power utilization in LLM inference clusters should not be very high. Our deductions align with data from production LLM clusters, revealing that inference workloads offer substantial headroom for power oversubscription. However, the stringent set of telemetry and controls that GPUs offer in a virtualized environment make it challenging to build a reliable and robust power management framework.We leverage the insights from our characterization to identify opportunities for better power management. As a detailed use case, we propose a new framework called POLCA, which enables power oversubscription in LLM inference clouds. POLCA is robust, reliable, and readily deployable. Using open-source models to replicate the power patterns observed in production, we simulate POLCA and demonstrate that we can deploy 30% more servers in existing clusters with minimal performance loss. 

To mitigate climate change, we must reduce carbon emissions from hyperscale cloud computing. We find that cloud compute servers cause the majority of emissions in a general-purpose cloud. Thus, we motivate designing carbon-efficient compute server SKUs, or GreenSKUs, using recently-available low-carbon server components. To this end, we design and build three GreenSKUs using low-carbon components, such as energy-efficient CPUs, reused old DRAM via CXL, and reused old SSDs. We detail several challenges that limit GreenSKUs, carbon savings at scale and may prevent their adoption by cloud providers. To address these challenges, we develop a novel methodology and associated framework, GSF (GreenSKU Framework), that enables a cloud provider to systematically evaluate a GreenSKU’s carbon savings at scale. We implement GSF within Microsoft Azure’s production constraints to evaluate our three GreenSKUs’ carbon savings. Using GSF, we show that our most carbon-efficient GreenSKU reduces emissions per core by 28% compared to currently-deployed cloud servers. When designing GreenSKUs to meet applications’ performance requirements, we reduce emissions by 15%. When incorporating overall data center overheads, our GreenSKU reduces Azure’s net cloud emissions by 8%. 

When a moderately intense, few-picosecond-long laser pulse ionizes gas to produce an underdense plasma column, a linear relativistic plasma wave or wake can be excited by the self-modulation instability that may prove useful for multi-bunch acceleration of externally injected electrons or positrons to high energies in a short distance. At the same time, due to the anisotropic temperature distributions of the ionized plasma electrons, the Weibel instability can self-generate magnetic fields throughout such a plasma on a few picoseconds timescale that can persist even longer than the lifetime of the wake. In the present paper, we first show using simulations that both these effects do indeed co-exist in space and time in the plasma. Using our simulations, we make preliminary estimates of the contribution to the transverse emittance growth of an externally injected beam due to the Weibel magnetic fields in a few-millimeter-long plasma. We then present the results of an experiment that has allowed us to measure the spatiotemporal evolution of the magnetic fields using an ultrashort relativistic electron probe beam. Both the topology and the lifetime of the Weibel instability induced magnetic fields in the experiment are in reasonable agreement with the fields induced by the Weibel instability in the simulations. 

Optical Thomson scattering is now a mature diagnostic tool for precisely measuring local plasma density and temperature. These measurements typically take advantage of a simplified analytical model of the scattered spectrum, which is built upon the assumption that each plasma species is in thermal equilibrium. However, this assumption fails for most laboratory plasmas of interest, which are often produced through high field ionization of atoms via ultrashort laser pulses and vulnerable to several kinetic instabilities. While it is possible to analytically model the Thomson scattered spectrum for some non-Maxwellian distribution functions, it is often not practical to do so for laboratory plasmas with highly complex and unstable distribution functions. We present a new method for predicting the Thomson scattered spectrum from any plasma directly from fully kinetic particle-in-cell simulations. This approach allows us to model the contributions of kinetic instabilities to the Thomson spectrum that aren’t taken into account in Maxwellian theory. We demonstrate this method’s capability to capture nonthermal features in the Thomson spectrum by simulating a simple bumpon- tail plasma as well as a more complex laser-ionized plasma. The versatility of this approach makes it an effective aid in the experimental design of Thomson diagnostics to directly characterize kinetic instabilities in laboratory plasmas. Index Terms—plasma measurement, low-temperature plasmas, plasma diagnostics, plasma simulation, plasma stability, plasma density, plasma temperature 

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. 

Absolute density measurements of low-ionization-degree or low-density plasmas ionized by lasers are very important for understanding strong-field physics, atmospheric propagation of intense laser pulses, Lidar etc. A cross-polarized common-path temporal interferometer using balanced detection was developed for measuring plasma density with a sensitivity of ∼0.6 mrad, equivalent to a plasma density-length product of ∼2.6 × 10¹³cm⁻²if using an 800 nm probe laser. By using this interferometer, we have investigated strong-field ionization yield versus intensity for various noble gases (Ar, Kr, and Xe) using 800 nm, 55 fs laser pulses with both linear (LP) and circular (CP) polarization. The experimental results were compared to the theoretical models of Ammosov-Delone-Krainov (ADK) and Perelomov-Popov-Terent’ev (PPT). We find that the measured phase change induced by plasma formation can be explained by the ADK theory in the adiabatic tunneling ionization regime, while PPT model can be applied to all different regimes. We have also measured the photoionization and fractional photodissociation of molecular (MO) hydrogen. By comparing our experimental results with PPT and MO-PPT models, we have determined the likely ionization pathways when using three different pump laser wavelengths of 800 nm, 400 nm, and 267 nm. 

The origin of the seed magnetic field that is amplified by the galactic dynamo is an open question in plasma astrophysics. Aside from primordial sources and the Biermann battery mechanism, plasma instabilities have also been proposed as a possible source of seed magnetic fields. Among them, thermal Weibel instability driven by temperature anisotropy has attracted broad interests due to its ubiquity in both laboratory and astrophysical plasmas. However, this instability has been challenging to measure in a stationary terrestrial plasma because of the difficulty in preparing such a velocity distribution. Here, we use picosecond laser ionization of hydrogen gas to initialize such an electron distribution function. We record the 2D evolution of the magnetic field associated with the Weibel instability by imaging the deflections of a relativistic electron beam with a picosecond temporal duration and show that the measured k -resolved growth rates of the instability validate kinetic theory. Concurrently, self-organization of microscopic plasma currents is observed to amplify the current modulation magnitude that converts up to ~1% of the plasma thermal energy into magnetic energy, thus supporting the notion that the magnetic field induced by the Weibel instability may be able to provide a seed for the galactic dynamo. 

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 (carbon- content, 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 character- ized, 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 dif- ferent sources of variable capacity. Finally, we show variable resource capacity presents new scheduling challenges. We show how variation can cause significant performance degra- dation 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.