v0.1.4

v0.1.4 focuses on output analysis

Implemented B/S architecture. Provide a completed development framework basis for future work.

Build ScalerRun, Analyzerserver and webvisualizer.

Finished view:
Config: Config editor and execution controller
Analysis: Time and counting analysis, 3 thread aggregation mode, multi-process view, multi-thread view, symbol view
About: About scaler

Improved binary compatibility, can run on Ubuntu 18.04, Cent OS 8 and Ubuntu 20.04

v0.1.3

v0.1.3 focuses on compatibility

Major improvements:

• Support both .plt and .plt.sec
• Tested on multiple machines with different linux distributions
• Tested with multiple real applications (TensorRT, Firefox .etc)
• Fixed bugs found during tests.

v0.1.2

v0.1.2 focuses on performance improvements.

Prehook+Afterhook+Counting+Timing

Parsec:
Overhead: 25.67%
Memory Overhead: 31.47%

SPEC:
Overhead: 4%

Prehook+Counting

Parsec:
Overhead: 5.22%
Memory Overhead: 20.99%

For details, check #11

Major improvements:

• Rm redundant code
• Optimize TLS
• Support different modes.
• Support prehook only
• Rm non global symbols hook

v0.1.1

v0.1.0 version has two functions accidentally removed by two last commits.
It also unhooks more jmp and exception functions.

This should be a new baseline before performance tuning.

See issue #4 and #5

v0.1.0

v0.1.0 version features a fully working libScalerHook assembly version and a visualizer.
Next, we first need to improve the performance of libScalerhook and then implement tools to analyze profiling data.

LibScalerhook:

• Implemented 3 versions (Assembly, Ptrace-Breakpoint, Breakpoint)
  • Assembly: Fully working
  • Ptrace-Breakpoint: Fully working (But multiple recent bug fixes in the assembly version may affect this version. Minor adjustments are necessary to make it work again. This version may be abandoned or used for comparison only because of too large overhead. The last working commit was d7a79bf.
  • Breakpoint: Only with prehook. According to the test, it's faster than Ptrace-Breakpoint but is still considered as too slow even if only prehook is implemented. This version may be abandoned or used for comparison only because of too large overhead. The last working commit was 1526205.

LibAnalyzer:

• Implemented visualization tools from both perf and libscalerhook data. Data format is defined in libScalerhookOutputFormat.md