Low Latency Systems and Communication Control
English:
How to design low-latency systems: controlling communication, Disruptor, Aeron, and the trade-offs between speed and architecture.
Highload
Highload on ThecoreGrid focuses on designing and operating systems that handle massive scale, traffic, and data under strict reliability requirements.
We explore architectures and patterns for horizontal scaling, load distribution, fault tolerance, and performance optimization in distributed environments. Topics include sharding, replication, caching strategies, queueing systems, backpressure handling, and latency reduction under peak load. We analyze real-world trade-offs between consistency, availability, and cost, along with failure scenarios and recovery strategies. Content is grounded in BigTech practices, including incident post-mortems and lessons from operating systems at global scale. You’ll find deep dives into infrastructure behavior, traffic management, autoscaling, and resilience engineering. Instead of simplified guides, the Highload tag delivers practical engineering insights for backend engineers, architects, platform teams, and SREs responsible for building and maintaining systems that must perform reliably under extreme demand.
CPU-free LLM inference without CPU involvement
CPU-free LLM inference: how to remove the CPU from the critical path and stabilize latency in LLM serving architectures.
KV cache optimization for multi-LoRA agents
KV cache optimization in multi-LoRA serving: how ForkKV reduces memory consumption and increases throughput of LLM inference.
Platform Program split at Uber under growth pressure
Platform Program split became a key step for Uber when the growth of the team began to hinder development. This decision changed both the architecture and the organization simultaneously. The problem manifested not at the code level, but at the level of team interaction. When Uber’s engineering organization grew to about 100 people, the division … Read more
P2P Model Distribution in Kubernetes Without Bottlenecks
P2P model distribution in Kubernetes with Dragonfly: how to reduce traffic to the origin and accelerate the delivery of large models from Hugging Face and ModelScope.
Symbolic execution for BPF malware analysis
Symbolic execution simplifies the analysis of BPF malware and eliminates a bottleneck in reverse engineering. This approach allows for the automatic reconstruction of “magic” packets to trigger backdoors. The problem does not manifest immediately — until the analysis of BPF malware encounters the complexity of the filters themselves. The classic Berkeley Packet Filter operates as … Read more
DWDP for LLM Inference Without Inter-GPU Synchronization
How DWDP optimizes LLM inference by eliminating inter-GPU synchronization and increasing throughput in multi-GPU systems.
Topology-preserving compression at GB per second
Topology-preserving compression without sacrificing speed: how EXaCTz achieves GB/s throughput while preserving the contour tree and extremum graph.
Network slicing with trust constraints without latency loss
Online network slicing with trust constraints: how the Path–Link model reduces latency and accelerates VNF placement in multi-domain infrastructure.
Reverse Address Translation in multi-GPU systems
How Reverse Address Translation affects latency in multi-GPU systems and why TLB misses hinder All-to-All operations in ML workloads.