Linux VDS Running Slow Despite Normal CPU? Here Is a Diagnostic Runbook

·31 views

A new technical runbook published on DEV Community addresses a common but misleading scenario where a Linux virtual server appears healthy by CPU metrics yet delivers slow or intermittent performance to users. The guide argues that CPU utilization alone cannot reveal whether a system is stalling on storage I/O, memory reclaim, network issues, cgroup throttling, or hypervisor steal time. It outlines a structured, read-only-first approach to incident diagnosis, starting with capturing a precise identity bundle including timestamps, hostname, and symptom description before collecting any metrics. The runbook walks through tools such as vmstat, mpstat, pidstat, and Linux Pressure Stall Information to isolate whether bottlenecks are guest-side or originate upstream in the infrastructure. It also cautions engineers to avoid running load-generation tools like fio or iperf3 on production systems without an approved maintenance window and explicit rate limits.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Working citation links do not prove quotes are accurate, lab tests find

A DEV Community lab tested whether AI citation checks can rely on links that open successfully. In 60 simulated answers, every cited local file existed, yet the check had to confirm the quoted words themselves. Exact matching of the quote as a contiguous run of tokens in a stored source caught all wrong-document, merged, and reworded quotes. Simple normalization was needed because raw byte comparison wrongly rejected a third of genuine quotes. The author warns that normalization itself can erase meaning, such as signs, decimal points, and superscripts, so it should preserve numbers as written.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Governed retrieval for duplicate healthcare appointment evidence keeps source citations intact

The article argues that near-duplicate appointment records should not be discarded during retrieval, since collapsing them early removes the audit trail a reviewer needs. It recommends staged retrieval where each source keeps a stable document identifier and revision, and where citation tracing acts as a gate before any answer is shown. Possible duplicates are grouped after retrieval as an application-level decision, with both original records preserved until enough source context exists to merge or send them for review. The author also presents Infrai as a plain-HTTP retrieval option with a single key and bill across many routes, while stressing that the application itself must own evidence identity and citation policy.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Atomic claims stop AI agents from repeating each other's jobs, simulation finds

A simulation run on 2026-11-13 showed that multi-agent systems can waste work when agents check a shared task list and then mark a job as taken in separate steps. Uncoordinated workers duplicated about 15 of 30 jobs per trial on average, while an atomic claim that combines the check and the mark produced no duplicates across 100 trials. The approach pairs each claim with a time-limited lease, so a job held by a crashed agent is released and finished by another within about 650 virtual milliseconds on average. A single dispatcher also eliminated duplicates in the test. The author notes that leases reduce duplicates but cannot fully rule them out, since an agent may pause and lose its claim.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why PCB autorouters fall short and how one handles unfinished boards

An autorouter converts schematic connections into copper across chosen layers while respecting design rules, but many tools fall short in practice. The article argues that a useful router should work across many layers, rework routing around difficult nets, and handle impedance during routing rather than afterward. It also stresses reporting every unrouted net by name instead of a vague completion figure. Using a four-layer RP2040 board as an example, the author describes a run that left nine nets open and refused to export Gerbers, which they present as the proper way to fail. Finally, the piece notes that component placement and dense fine-pitch parts often cause routing failures more than the router itself.

0 comments Read more at DEV Community