System and method for software/hardware coordinated adaptive performance monitoring

A method for adaptive performance monitoring to manage memory cache of a processing system, the method comprising:

• polling, at a current polling instance, processor counters to manage last level cache (LLC);
• computing performance metrics according to the processor counters;
• determining whether to enable adaptation of polling processor counters according to the performance metrics;
• upon determining to enable adaptation, adjusting a time interval for a next polling instance according to the performance metrics;
• and
• resetting the time interval for next polling to a pre-determined value and the performance metrics upon determining unstable system conditions beyond at least one a threshold of instructions per cycle and a cache hit ratio threshold,
• wherein the processor counters include an instruction count and a cycle count per each processor core and a hit count and a miss count of LLC hits between last two polling instances,
• wherein the performance metrics include executed instructions per cycle (IPC) computed as a sum of ratios of the instruction count to the cycle count of all processor cores, and a cache hit ratio (CHR) computed as a ratio of the hit count to a sum of the hit count and the miss count, and
• wherein the performance metrics further include a cost function computed as a weighted sum of an IPC difference and a CHR difference between last two polling instances.

An apparatus configured for adaptive performance monitoring to manage memory cache of the apparatus, the apparatus comprising:

• a multi-core processor;
• a multi-level cache;
• and
• a computer readable storage medium storing programming for execution by the multi-core processor, the programming including instructions to:
• poll performance counters of the multi-core processor at each of a plurality of subsequent instances;
• manage the multi-level cache using the performance counters;
• adjust a time interval for polling the performance counters at a next instance according to the performance counters captured at a last instance;
• and
• reset the time interval to a predetermined value upon determining the performance counters reflect unstable system conditions according to at least one of a number of instructions per cycle (IPC) and a cache hit ratio (CHR), wherein the instruction to adjust the time interval comprises to estimate the time interval based on a current polling instance and a previous polling instance.