y_uuki
Re:
on Monday August 03, 2026 @06:23PM (#49162244)
Grafana app plugin for monitoring Slurm jobs on GPU clusters. View per-job GPU, CPU, memory, and network metrics with automatic time range and node filtering.
Re:An eBPF-Based Prometheus Exporter for Lustre Client
on Monday August 03, 2026 @05:13AM (#49153197)
lustre-ebpf-exporter measures Lustre client activity continuously on the client node. It uses llite as the primary observation plane for user-facing access activity and PtlRPC as the secondary observation plane for client-internal wait behavior.
Prometheus Exporter for RDMA (RoCE) NIC Statistics in Linux
A leader election daemon using Consul Session and KV Lock for VIP failover
Re:tcpulse: A TCP/UDP load generator that provides fine-grained, flow-...
on Tuesday June 10, 2025 @01:58AM (#44233079)
Currently tcpulse only supports a `pingpong` mode, so true one-way (“diode”) transfers aren’t available out of the box. That said, the design is simple enough that you could introduce an option or wrapper to lock the flow in one direction and implement a diode mode. Happy to review a PR if someone wants to add it
Re:tcpulse: A TCP/UDP load generator that provides fine-grained, flow-...
on Monday June 09, 2025 @11:33AM (#44225535)
iperf3 is a link “speedometer” – spin it up between two hosts, crank -P or -u -b, and it tells you max TCP/UDP throughput (and jitter/loss if you like).
tcpulse is a fine-grained traffic “microscope” – you dial exact CPS or concurrent sockets, spray dozens of targets from one client, and get p90/p95/p99 latencies per flow.
Use iperf3 for a quick bandwidth check; use tcpulse when you need repeatable, controlled connection patterns and detailed latency stats across many backends.
Re:tcpulse: A TCP/UDP load generator that provides fine-grained, flow-...
on Monday June 09, 2025 @11:26AM (#44225465)
Thank you!
No, tcpulse only measures and prints out metrics from the client side, not the server side. However, since multiple clients each output their own measurement results, this effectively results in performance metrics being output for each peer.
tcpulse: A TCP/UDP load generator that provides fine-grained, flow-level control
Re:The AIOps library for removing automatically failure-unrelated metrics
on Monday August 26, 2024 @05:07AM (#41355282)
MetricSifter is a feature reduction framework designed to accurately identify anomalous metrics caused by faults for enhancing fault localization. Our key insight is that the change point times inside the failure duration are close to each other for the failure-related metrics. MetricSifter detects change points per metric, localizes the time frame with the highest change point density, and excludes metrics with no change points in that time frame. The offline change point detection is implemented by ruptures, and the segmentation of the detected change points is based on kernel density estimation (KDE).