Transparent performance inference of whole software layers and context-sensitive performance debugging
Methods and systems for performance inference include inferring an internal application status based on a unified call stack trace that includes both user and kernel information by inferring user function instances. A calling context encoding is generated that includes information regarding function calling paths. Application performance is analyzed based on the encoded calling contexts. The analysis includes performing a top-down latency breakdown and ranking calling contexts according to how costly each function calling path is.
1. A method for performance inference, comprising:
inferring an internal application status based on a unified call stack trace that includes both user and kernel information by inferring latencies for individual function calls in user function instances;
generating a calling context encoding that includes information regarding function calling paths and represents each calling context as a distinct integer; and
analyzing application performance based on the encoded calling contexts, comprising:
performing a top-down latency breakdown, comprising:
estimating a total running time of each function in a stack trace; and
excluding running times attributable to later-starting, concurrently running functions to determine a latency attributable to each function;
annotating the calling context encoding with inferred performance information; and
ranking calling contexts according to how costly each function calling path is.
2. The method of claim 1 , wherein estimating a total running time comprises finding a difference between a first stack trace in which a function appears and a first stack trace in which the function no longer appears.
3. The method of claim 1 , wherein estimating a total running time comprises finding a difference between a first stack trace in which a function appears and a last uninterrupted stack trace in which the function appears.
4. The method of claim 1 , wherein ranking calling contexts comprises ranking according to total cost for each function calling path.
5. The method of claim 1 , wherein ranking calling contexts comprises ranking according to a difference in latency between two input workloads.
6. A system for performance inference, comprising:
an inference module comprising a hardware processor configured to infer an internal application status based on a unified call stack trace that includes both user and kernel information by inferring latencies for individual function calls in user function instances, to generate a calling context encoding that includes information regarding function calling paths, and to annotate the calling context encoding with inferred performance information; and
a performance analysis module configured to analyze application performance based on the encoded calling contexts by estimating a total running time of each function in a stack trace and excluding running times attributable to later-starting, concurrently running functions to determine a latency attributable to each function in a top-down latency breakdown and by ranking calling contexts according to how costly each function calling path is.
7. The system of claim 6 , wherein the performance analysis module is further configured to find a difference between a first stack trace in which a function appears and a first stack trace in which the function no longer appears.
8. The system of claim 6 , wherein the performance analysis module is further configured to find a difference between a first stack trace in which a function appears and a last uninterrupted stack trace in which the function appears.
9. The system of claim 6 , wherein the performance analysis module is further configured to rank calling contexts according to total cost for each function calling path.
10. The system of claim 6 , wherein the performance analysis module is further configured to rank calling contexts according to a difference in latency between two input workloads.