SAND: A New Programming Abstraction for Video-based Deep Learning

solve heavy preprocessing problem with abstraction & planning + priority scheduling to reuse

Tiga: Accelerating Geo-Distributed Transactions with Synchronized Clocks

Sleeping with One Eye Open:Fast, Sustainable Storage with Sandman

cache_ext: Customizing the Page Cache with eBPF

OS page cache evicition policy: only 1 -> supports multiple policy, FLEXIBILITY

A large scale analysis of hundreds of in-memory cache clusters at Twitter

5 important facts about in-memory caching

Jenga: Effective Memory Management for Serving LLM with Heterogeneity

a memory allocation framework for managing heterogeneous embeddings by leveraging layer properties

PrefillOnly: An Inference Engine for Prefill-only Workloads in Large Language Model Applications

We only need to store 1 layer. Let's not do unnecessary things and UTILIZE this condition.

IC-Cache: Efficient Large Language Model Serving via In-context Caching

In-Context Caching system/ leverage caching, let small LLM serves like a giant.

COpter: Efficient Large-Scale Resource-Allocation via Continual Optimization

remove overhead of round-based resource allocation -> sequence of interconnected problems

Mitigating Application Resource Overload with Targeted Task Cancellation

Eliminate Head-Of-Line Blocking request to solve resource overload problem.

Kinetic Modeling of Data Eviction in Cache

AET — composable, linear-time Miss Ratio Curve profiling via average eviction time sampling

Sailor: Automating Distributed Training over Dynamic, Heterogeneous, and Geo-distributed Clusters

Co-optimize resource allocation and parallelization plan for heterogeneous GPU training with accurate simulation

Robust LLM Training Infrastructure at ByteDance

Automated failure diagnosis and recovery for large-scale LLM training — minimize unproductive time across 16K+ GPUs

Spirit: Fair Allocation of Interdependent Resources in Remote Memory Systems

Fair allocation when resources are interdependent — remote memory bandwidth, capacity, and compute interact

Oasis: Pooling PCIe Devices Over CXL to Boost Utilization

Use CXL shared memory as a message channel to pool PCIe devices (NICs, SSDs) across hosts in a CXL pod

Disentangling the Dual Role of NIC Receive Rings

Split Rx ring into allocation (Ax) and reception (Bx) rings → reduce I/O working set, improve throughput up to 37%

Criticality-Aware Instruction-Centric Bandwidth Partitioning for Data Center Applications

Static partitioning can't react quickly to load changes → use fine-grained control signals and core allocation instead

Extending Applications Safely and Efficiently

EIM model abstracts extension resources for fine-grained safety/interconnectedness tradeoffs; bpftime enforces it efficiently

Tile Size Selection Using Cache Organization and Data Layout

Select tile sizes that fit data working sets in cache, accounting for layout and associativity

SOCK: Rapid Task Provisioning with Serverless-Optimized Containers

Lean containers + generalized Zygote provisioning + three-tier package-aware caching → 45× cold-start speedup

Shared Address Translation Revisited

Share page tables across processes for shared libraries → reduce TLB overhead and page faults on Android

SEUSS: Skip Redundant Paths to Make Serverless Fast

Deploy serverless functions from unikernel snapshots to skip boot, runtime init, and import overhead — 51× throughput improvement

Serverless in the Wild: Characterizing and Optimizing the Serverless Workload at a Large Cloud Provider

Characterize FaaS workloads → propose per-function keep-alive and pre-warm policies to cut cold starts

Replayable Execution Optimized for Page Sharing for a Managed Runtime Environment

Checkpoint with GC-compacted heap → share pages across containers → speedy restoration with fewer page faults

Benchmarking, Analysis, and Optimization of Serverless Function Snapshots

Functions access stable working sets across invocations → prefetch from disk to cut cold-start latency by 3.7×

Prebaking Functions to Warm the Serverless Cold Start

The timing of the snapshot determines cold-start latency — prebake at the right execution point

Reducing Minor Page Fault Overheads through Enhanced Page Walker

Hardware-software co-design offloads minor page fault critical path → 33× latency improvement, 6.6% runtime improvement

MEGA: Overcoming Traditional Problems with OS Huge Page Management

Analyze and address the fundamental problems with Linux huge page management — fragmentation, bloat, fault latency, non-swappability

Coordinated and Efficient Huge Page Management with Ingens

Treat memory contiguity as a first-class resource; track utilization and access frequency for principled huge page management

Memory Efficient Fork-based Checkpointing Mechanism for In-Memory Database Systems

Fork-based checkpointing with efficient copy-on-write management to minimize memory overhead for in-memory databases

FlashCube: Fast Provisioning of Serverless Functions with Streamlined Container Runtimes

Streamline container runtimes to eliminate unnecessary initialization overhead for fast serverless provisioning

Parallelizing Packet Processing in Container Overlay Networks

Serialized softirqs on a single core bottleneck overlay networks → pipeline them across multiple cores with Falcon

FaaSNet: Scalable and Fast Provisioning of Custom Serverless Container Runtimes at Alibaba Cloud Function Compute

Decentralize container provisioning across host VMs in function-tree structures to eliminate cold-start bottleneck

Architectural Implications of Function-as-a-Service Computing

FaaS containerization brings up to 20× slowdown vs. native; cold start can exceed 10× a function's execution time

Cold Start Influencing Factors in Function as a Service

Programming language, package size, and memory/CPU settings significantly affect FaaS cold-start latency

Caladan: Mitigating Interference at Microsecond Timescales

Dedicated scheduler core + fast core allocation reacts to interference in microseconds — no slow hardware partitioning