How and When You Should Measure CPU Overhead of eBPF Programs Bryce Kahle, Datadog October 28, 2020 Why should I profile eBPF programs? CI variance tracking Tools kernel.bpf_stats_enabled kernel cases: – Benchmarking + CI/CD – Sampling profiler in production How does it work? – Adds ~20ns of overhead per run Two ways to enable kernel eBPF stats sysctl procfs Three ways to access kernel eBPF stats

and reliability secCrypto secPaver secGear Simplified O&M and development A-Ops SysCare CPDS CPU GPU Optimal performance for a single scenario Multi-scenario capability collaboration and sharing automatically generates module files. • HPCRunner implements one-click compilation and operation, CPU/GPU performance profiling, and benchmarking based on HPC configurations. • All configurations are for optimization based on specific service characteristics and requirements. For example, specific CPU or device binding policies can be implemented. To meet the requirements of these applications, hwloc

enhances server and cloud computing features, and incorporates key technologies including cloud-native CPU scheduling algorithms for hybrid service deployments and KubeOS for containers. As an OS platform suits hybrid deployments of online and offline cloud services. Its innovative CPU scheduling algorithm ensures real-time CPU preemption and jitter suppression for online services. Additionally, its innovative Docker+QEMU solution, the iSulad+shimv2+StratoVirt secure container solution reduces the memory overhead and boot time by 40%. • Dual-plane deployment tool eggo: OSs can be installed with one click for

- Flexibility Why BPF? - Flexibility - Low overhead Why BPF? - Flexibility - Low overhead - Continuous profiling Why BPF? - Flexibility - Low overhead - Continuous profiling - No modifications of - Trace complex Ruby programs execution rbperf – on-CPU profiling - $ rbperf record --pid=124 cpu - $ rbperf report [...] rbperf – Rails on-CPU profile rbperf – tracing write(2) calls - $ rbperf driver program - Make the OSS version awesome - Better documentation (including how to measure overhead) - Add more output formats - Open source GDB / drgn helper - Other tools? - Containers support

relying on eBPF, all visibility is programmable and allows for a dynamic approach that minimizes overhead while providing deep and detailed visibility as required by users. Hubble has been created and specifically in the Linux kernel’s socket layer (e.g. at TCP connect time) such that per-packet NAT operations overhead can be avoided in lower layers. Bandwidth Management Cilium implements bandwidth management through If you are running in an environment with more than 250 nodes, 5k pods, or if you observe a high overhead in state propagation caused by Kubernetes events. If you do not want Cilium to store state in Kubernetes

relying on eBPF, all visibility is programmable and allows for a dynamic approach that minimizes overhead while providing deep and detailed visibility as required by users. Hubble has been created and specifically in the Linux kernel’s socket layer (e.g. at TCP connect time) such that per-packet NAT operations overhead can be avoided in lower layers. Bandwidth Management Cilium implements bandwidth management through of reasons when to use a kvstore: If you are running in an environment where you observe a high overhead in state propagation caused by Kubernetes events. If you do not want Cilium to store state in Kubernetes

relying on eBPF, all visibility is programmable and allows for a dynamic approach that minimizes overhead while providing deep and detailed visibility as required by users. Hubble has been created and specifically in the Linux kernel’s socket layer (e.g. at TCP connect time) such that per-packet NAT operations overhead can be avoided in lower layers. Bandwidth Management Cilium implements bandwidth management through If you are running in an environment with more than 250 nodes, 5k pods, or if you observe a high overhead in state propagation caused by Kubernetes events. If you do not want Cilium to store state in Kubernetes

relying on BPF, all visibility is programmable and allows for a dynamic approach that minimizes overhead while providing deep and detailed visibility as required by users. Hubble has been created and specifically If you are running in an environment with more than 250 nodes, 5k pods, or if you observe a high overhead in state propagation caused by Kubernetes events. If you do not want Cilium to store state in Kubernetes number of open connections. Thus, clients are encouraged to cache their connections rather than the overhead of reopening TCP connections every time they need to store or retrieve data. Multiple clients can

platform Compass-CI Toolchain OpenStack Kubernetes Kylin HA Cluster scheduling and management CPU: x86, ARM, RISC-V GPU NPU Chips Apps Virtualization Containers QEMU Docker libvirt Virtualization/ least 4 GB Hard drives At least 20 GB Item Configuration Requirement Architecture AArch64, x86_64 CPU 2 CPUs Memory At least 4 GB Hard drives At least 20 GB 10 11 openEuler WHITE PAPER openEuler utilization, and fewer invalid migrations. 2. Enhanced CPU isolation: The isolation of interrupts and unbound kthreads further enhances the isolation of CPU cores and minimizes mutual interference between

number of open connec�ons. Thus, clients are encouraged to cache their connec�ons rather than the overhead of reopening TCP connec�ons every �me they need to store or retrieve data. Mul�ple clients can benefit in-kernel verifier ensures that BPF programs are safe to run and a JIT compiler converts the bytecode to CPU architecture specific instruc�ons for na�ve execu�on efficiency. BPF programs can be run at various between 10 seconds and 30 minutes or 12 hours for LRU based maps. This should automa�cally op�mize CPU consump�on as much as possible while keeping the connec�on tracking table u�liza�on below 25%. If needed