pcm-iio Tool outputs "Server CPU is required for this tool! Program aborted". Is there a way I can monitor my PCIe link bandwidth using one of the PCM utilities on client CPU?
Answer: The "IO" metric in pcm utility is the closest capability to monitor I/O PCIe traffic on client CPUs.
PCM reports "ERROR: QPI LL monitoring device (...) is missing". How to fix it?
Answer: It is likely a BIOS issue. See details here
Does PCM work inside a virtual machine?
Answer: PCM works inside virtual machines which support vPMU (with a limited set of metrics supported by vPMU). For example on AWS instances this is indicated by the presence of arch_perfmon flag: https://instaguide.io/ . Enabling vPMU in hypervisors is documented in Hardware Event-Based Sampling section.
Does PCM work inside a docker container?
Answer: yes, it does. An example of how to run PCM inside a docker container is located here. The recipe works also for other PCM utilities besides pcm-sensor-server.
pcm-power reports "Unsupported processor model". Can you add support for pcm-power for my CPU?
Answer: most likely you have a client CPU which does not have required hardware performance monitoring units. PCM-power can not work without them.
pcm-memory reports that the CPU is not supported. Can you add support for pcm-memory for my CPU?
Answer: most likely you have a client CPU which does not have required hardware performance monitoring units. PCM-memory can not work without them.
Can PCM be used for measuring energy, CPU cycles, etc for a particular process or does it measure for the system as a whole?
Answer: PCM supports measurement for the whole system, per processor, per physical or per logical core. If you need monitoring per-process or user per-thread you can pin your process and/or thread to certain cores and read PCM data for these cores. But keep in mind that the OS can also schedule other processes or threads on this core and this may disturb your measurements. For a precise per-process or per-thread measurement the Intel VTune profiler or Linux perf profiler should be used.
PCM reports
opening /dev/mem failed: errno is 1 (Operation not permitted)
Can not read memory controller counter information from PCI configuration space. Access to memory bandwidth counters is not possible.
You must be root to access these SandyBridge/IvyBridge/Haswell counters in PCM.
Secure Boot detected. Using Linux perf for uncore PMU programming.
How to fix it?
Answer: Linux disables access to /dev/mem because Secure Boot is enabled in the BIOS. Disable Secure Boot in the BIOS to enable memory controller statistics (e.g. memory read and write bandwidth).
PCM reports
Linux Perf: Error on programming ...
How to fix it?
Answer: It is an issue with the Linux kernel perf driver. As a workaround upgrade your Linux kernel to the latest available/possible or run PCM with its own programming logic:
export PCM_NO_PERF=1
pcm -r
If you are getting the error Starting MSR service failed with error 3 The system cannot find the path specified. try to uninstall the driver by running pcm --uninstallDriver and optionally reboot the system.
Is PCM supported on AWS instances
Answer:
Not all AWS instances allow users to collect CPU telemetry by exposing PMU to the user. The following instances can be used:
- Bare metal instances: allow collection of CPU metrics from both core (e.g. instructions per cycle, cache misses) and uncore (e.g. memory controller, UPI)
- Full-socket (single socket, two socket, etc) virtualized instances: e.g. m5d.12xlarge, m5.24xlarge, m5.12xlarge. Only core CPU metrics are exposed, and certain CPU performance events are forbidden (e.g. offcore response events, events collecting “any_thread” information). “arch_perfmon” flag in /proc/cpuinfo indicates if the core CPU metrics are exposed (example: https://instaguide.io/info.html?type=m5.12xlarge ). The mechanism of PMU virtualization is commonly known as vPMU.
pcm-pcie reports that the CPU is not supported: "Jaketown, Ivytown, Haswell, Broadwell-DE, Skylake, Icelake, Snowridge and Sapphirerapids Server CPU is required for this tool! Program aborted" Can you add support for pcm-pcie for my CPU?
Answer: most likely you have a client CPU which does not have required hardware performance monitoring units. pcm-pcie can not work without them.