arXiv subjects

Sebastian Burckhardt

Publications and source records attributed to Sebastian Burckhardt.

3 recordsLinked to original sources

arXiv2026

Tile-Level Activation Overlap for Efficient LLM Inference

SwiGLU is the dominant MLP activation in modern large language models, yet its intermediate tensor materialization costs 9-37% of MLP execution time. We present two complementary CUTLASS-based SM90 kernels that fuse SwiGLU into GeMM at the tile level. Kernel-1 overlaps Swish computation on the Gate accumulator with Up-tile loading using the Pingpong warp-specialized schedule; Kernel-2 interleaves SwiGLU with tile stores via a custom Epilogue Visitor Tree. Evaluated on Qwen-2.5 models (0.5B-72B) on NVIDIA H100, our kernels achieve up to 2.47x speedup over PyTorch, shifting workloads from memory-bound to compute-bound and reaching 79.5% peak BF16 utilization. We demonstrate that torch.compile cannot replicate this fusion (3-7x slower than our kernels), validating the need for hand-crafted tile-level design. Our fused kernels are also numerically superior, achieving zero mismatches compared to 4.5-11% for cuBLAS.

cs.DC↗

arXiv2021

Serverless Workflows with Durable Functions and Netherite

Serverless is an increasingly popular choice for service architects because it can provide elasticity and load-based billing with minimal developer effort. A common and important use case is to compose serverless functions and cloud storage into reliable workflows. However, existing solutions for authoring workflows provide a rudimentary experience compared to writing standard code in a modern programming language. Furthermore, executing workflows reliably in an elastic serverless environment poses significant performance challenges. To address these, we propose Durable Functions, a programming model for serverless workflows, and Netherite, a distributed execution engine to execute them efficiently. Workflows in Durable Functions are expressed as task-parallel code in a host language of choice. Internally, the workflows are translated to fine-grained stateful communicating processes, which are load-balanced over an elastic cluster. The main challenge is to minimize the cost of reliably persisting progress to storage while supporting elastic scale. Netherite solves this by introducing partitioning, recovery logs, asynchronous snapshots, and speculative communication. Our results show that Durable Functions simplifies the expression of complex workflows, and that Netherite achieves lower latency and higher throughput than the prevailing approaches for serverless workflows in Azure and AWS, by orders of magnitude in some cases.

cs.DC↗

arXiv2017

Consistency models with global operation sequencing and their composition (extended version)

Modern distributed systems often achieve availability and scalability by providing consistency guarantees about the data they manage weaker than linearizability. We consider a class of such consistency models that, despite this weakening, guarantee that clients eventually agree on a global sequence of operations, while seeing a subsequence of this final sequence at any given point of time. Examples of such models include the classical Total Store Order (TSO) and recently proposed dual TSO, Global Sequence Protocol (GSP) and Ordered Sequential Consistency. We define a unified model, called Global Sequence Consistency (GSC), that has the above models as its special cases, and investigate its key properties. First, we propose a condition under which multiple objects each satisfying GSC can be composed so that the whole set of objects satisfies GSC. Second, we prove an interesting relationship between special cases of GSC---GSP, TSO and dual TSO: we show that clients that do not communicate out-of-band cannot tell the difference between these models. To obtain these results, we propose a novel axiomatic specification of GSC and prove its equivalence to the operational definition of the model.

cs.DC↗