IBM (Canada)

Modern AI accelerators adopt dataflow architectures to achieve both high peak throughput (TOPS) and energy efficiency (TOPS/W). These designs feature wide datapaths and hierarchical scratchpad memories that supply dense compute arrays with high-bandwidth data access and extensive operand reuse. Complementing the compute–memory subsystem is a lightweight control path that orchestrates data movement, program loading, and register initialization. To reduce energy and area overheads, conventional processor features—such as instruction caches, execution stacks, and branch speculation—are deliberately omitted. While this streamlined design maximizes efficiency, it shifts a critical responsibility onto the compiler: transforming complex kernels into highly compact instruction streams that must fit entirely within the limited instruction buffers (IBUFFs) of the accelerator’s programmable units.In this paper, we introduce two novel compiler transformations—Loop Absorption (LA) and Loop Index Set Merging (LISM) for ultra compact code generation. Loop Absorption merges isomorphic sibling operations into a single loop body, while LISM unifies adjacent loops with similar bodies into a unified iteration space. Together, these complementary techniques eliminate redundant code patterns and produce compact hierarchical loop nests. We implement LA and LISM in the IBM Spyre compiler and evaluate them on diverse deep learning workloads including ResNet-50, Inception-v3, SSD, and BERT-Large. Across these models, our combined approach achieves a geometric mean compression of 1.48× over the baseline, enabling layers that previously exceeded IBUFF capacity to compile successfully.

AI Accelerators employ dataflow architectures to achieve impressive peak compute performance (TOPS) and processing efficiencies (TOPS/W). Typically, dataflow architectures use wide data-paths to connect off-chip memory to dense compute arrays (via hierarchy of on-chip memories/vector register files) for efficient data movement with reuse, as well as compute. Such architectures often possess an independent lightweight control-path for loading programs and initializing registers, and lack traditional architectural features like instruction cache and execution stacks. This poses a unique challenge to compiler requiring program generation of complex compute kernels to fit within an instruction buffer and allocating a limited set of scalar registers without support to spill to memory.This paper contributes a significant step towards spill-free compilation and proposes a Live range reduction optimization based on Affine expression propagation analysis. Our solution performs a global, compiler-directed analysis to model variable values as affine expressions of in-scope variables, enabling safe symbolic re-materialization of values at their use-sites leveraging near- by variables without introducing new operations. This shortens variable lifetimes, while significantly reducing register pressure without incurring program binary and execution overhead. The static nature and regular memory access patterns of AI applications make them well-suited for the proposed optimization. We demonstrate the effectiveness of the technique in the context of IBM Spyre accelerator and its compiler. Our results over a range of AI workloads spanning transformer and CNN models demonstrate spill-free code generation, with most of the workloads requiring less than 50% of the available registers.

In the rapidly evolving landscape of software development, software testing is critical for maintaining code quality and reducing defects. Effective test case prioritization employs techniques to identify defects early and ensure software quality. New avenues of research have explored using machine learning (ML) to automate the process, most current applications leverage a machine learning model using numerical features to prioritize the test cases. This study investigates the enhancement of this process by incorporating text-based features derived from git commit messages, which often include valuable information about code changes. Given that commit messages are often poorly written and inconsistent, we employ a large language model (LLM) to rewrite these messages based on code diffs, with the aim of improving the quality of their format and the information they contain. We then assess whether these refined commit messages, as an additional feature, contribute to better performance of the test case prioritization model. Our preliminary results indicate that the inclusion of LLM-enhanced commit messages leads to a noticeable improvement in prioritization effectiveness, suggesting a promising avenue for integrating natural language processing techniques in software testing workflows.

Query optimizers in RDBMSs search for execution plans expected to be optimal for given queries. They use parameter estimates, often inaccurate, and make assumptions that may not hold in practice. Consequently, they may select plans that are suboptimal at runtime, if estimates and assumptions are not valid. Therefore, they do not sufficiently support robust query optimization. Using ML to improve data systems has shown promising results for query optimization. Inspired by this, we propose Robust Query Optimizer, (Roq), a holistic framework based on a risk-aware learning approach. Roq includes a novel formalization of the notion of robustness in the context of query optimization and a principled approach for its quantification and measurement based on approximate probabilistic ML. It also includes novel strategies and algorithms for query plan evaluation and selection. Roq includes a novel learned cost model that is designed to predict the cost of query execution and the associated risks and performs query optimization accordingly. We demonstrate that Roq provides significant improvements in robust query optimization compared with the state-of-the-art.

Ensuring the reliability and availability of cloud services relies heavily on efficient root cause analysis (RCA) for cloud incidents. Traditionally, RCA involved labor-intensive manual investigations across various data sources, such as logs, metrics, and traces. Despite the increasing adoption of AI-driven assistants for RCA, their effectiveness for Site Reliability Engineers (SREs) is often hindered by low accuracy due to the inherent complexity of the task. This study introduces an on-call system powered by a large language model (LLM) designed to automate RCA processes for cloud incidents in practical, privacy-aware industrial settings. Our approach integrates incoming multimodal datasets (Logs, Metrics, Traces, Alerts-LMTA), aggregates critical runtime diagnostic information, including developer set alerts with precise offset values and thresholds, and utilizes LMTA data to predict the root cause category of incidents. We evaluated the effectiveness of our approach using two open-source, real-world-like datasets, demonstrating that the proposed LLM based approach can achieve an RCA accuracy of up to 97%.

Java-based applications rely on just-in-time (JIT) compilation to convert bytecode into machine code at run-time to improve throughput. However, JIT compilation is achieved at the cost of additional CPU cycles and memory. In the context of modern cloud, where applications have limited computing resources due to containerization, JIT compilation may potentially become a pain-point. The Semeru Cloud Compiler (SCC) is a promising solution that offloads JIT compilation to a remote cloud compiler to reduce the local overhead of JIT compilation in containers. In this paper, we outline novel research directions to further leverage the advantages of remote cloud compilation by extending the SCC to achieve two goals: (i) reduce application start-up time and (ii) improve JIT compilation decision making. To reduce application start-up time, the proposed enhancement is to extract unused or unlikely to be used Java class files from container images. The SCC can still provide the extracted class files to the client JVM if any turn out to be needed. Our preliminary experiments show that this approach can potentially reduce container image sizes by up to 22.1% and start-up times by up to 9.2% for certain applications. To reduce JIT compilation overhead, the proposed enhancement is to perform dynamic analysis of the Java application using the additional resources on the server-side. In turn, the remote JIT compilation can make more informed decisions regarding which methods to compile and optimizations to apply. We are also developing a novel visualization tool designed to help JIT compiler developers identify pain points in current JIT compilation heuristics. Specifically, we have measured that up to 18.9% of JIT compilations for certain applications may be sub-optimal due to later method inlining, and we are investigating how the SCC could help to avoid these potentially wasted compile-time computations by using information learned from other JVM clients.

Resistance Temperature Detectors (RTD) were intricately microfabricated onto a lid, presenting a pioneering approach to improve Junction-to-Case thermal resistance measurements on electronic packages. Traditionally, those involve intrusive lid drilling and cold plate grooving to accommodate the thermocouples integration as stated in JEDEC JESD51 standards. In lieu of this intrusive practice, microfabricated RTDs were adopted. A configuration of twelve RTDs was carefully distributed throughout the thermally active surface of an actual lid, offering a markedly less obtrusive measurement approach while enhancing temperature distribution and precision. Integration of these sensors to the lid surface, albeit inherently rough, was addressed through a spray coating planarization technique, effectively diminishing surface roughness without impinging on functionality. Results attest the functionality of RTDs across a temperature range from 20 to 105 degrees Celsius, exhibiting linear performance characteristics. The comprehensive analysis of sensor responses is detailed within this article.

Clinical discoveries largely depend on dedicated clinicians and scientists to identify and pursue unique and unusual clinical encounters with patients and communicate these through case reports and case series. This process has remained essentially unchanged throughout the history of modern medicine. However, these traditional methods are inefficient, especially considering the modern-day availability of health-related data and the sophistication of computer processing. Outlier analysis has been used in various fields to uncover unique observations, including fraud detection in finance and quality control in manufacturing. We propose that clinical discovery can be formulated as an outlier problem within an augmented intelligence framework to be implemented on any health-related data. Such an augmented intelligence approach would accelerate the identification and pursuit of clinical discoveries, advancing our medical knowledge and uncovering new therapies and management approaches. We define clinical discoveries as contextual outliers measured through an information-based approach and with a novelty-based root cause. Our augmented intelligence framework has five steps: define a patient population with a desired clinical outcome, build a predictive model, identify outliers through appropriate measures, investigate outliers through domain content experts, and generate scientific hypotheses. Recognizing that the field of obstetrics can particularly benefit from this approach, as it is traditionally neglected in commercial research, we conducted a systematic review to explore how outlier analysis is implemented in obstetric research. We identified two obstetrics-related studies that assessed outliers at an aggregate level for purposes outside of clinical discovery. Our findings indicate that using outlier analysis in clinical research in obstetrics and clinical research, in general, requires further development.