$ python3 pmwatch.py 0 # To watch DIMM0intel/intel-pmwatch: Intel® PMWatch (PersistentMemoryWatch) is a tool that monitors and reports the performance and health information metrics of the Intel® Optane™ DC Persistent Memory.
TODO:
- colorize
- multi-dimm watch
- collecting data from
pcm-memory - loads watch
The metrics in pmwatch (indicated by *) are as follows:
- *bytes_read/bytes_write: The actual amount of data read and written, calculated as (read_64B_ops_received - write_64B_ops_received) * 64 for reads and write_64B_ops_received * 64 or media_write_ops * 256 for writes.
- *read_hit_ratio/write_hit_ratio: If write > 0.75 or read is close to 1, it indicates a buffer hit. If read or write = 0.75, it indicates a sequential access pattern.
- *media_read_ops/media_write_ops: The number of reads and writes on the media, measured in units of 256B. Calculated using the received values, where
media_read_ops = (read_64B_ops_received - write_64B_ops_received) / 4. - read_64B_ops_received/write_64B_ops_received: The number of operations on the media (including CPU-initiated and maintenance operations), measured in units of 64B. Note that this value may be greater than cpu_write_ops, as it includes operations initiated by the internal controller. The difference is around 12K during idle and increases during write-intensive operations, such as a single-threaded 256B write reaching ~123K.
- cpu_read_ops/cpu_write_ops: The number of read and write operations initiated by the memory controller, measured in units of 64B.
Therefore:
WA = media_write_ops * 4 / cpu_write_ops
RA = media_read_ops * 4 / cpu_read_ops
This script depends on PMWatch, which in turn depends on libipmctl. lingfenghsiang/pm_util depends on ipmctl show -performance, so the overhead should be similar.