Trident, a tool to use low level metrics derived from hardware
counters to understand Core, Memory and I/O utilisation and bottlenecks.
The collection of time series of these low level counters does not
induce significant overhead to the execution of the application.
The Understanding Performance team is investigating on a new node
characterisation tool, ¹Trident¹, that can look at various low level
metrics with respect to the Core, Memory and I/O. Trident uses a three
pronged approach to analysing node¹s utilisation and understand the
stress on different parts of the node based on the given job. Currently
core metrics such as memory bandwidth, core utilization, active processor
cycles, etc., are being collected. Interpretation of this data is often
non intuitive. The tool preprocesses the data to make the data usable by
developers and site managers without the need of in-depths expertise of
CPU and systems architecture details.