We introduce a hardware-based methodology for performing workload execution forensics in microprocessors. More specifically, we discuss the on-chip instrumentation required for capturing the operational profile of the Translation Lookaside Buffer (TLB), as well as an off-line machine learning approach which uses this information to identify the executed processes and reconstruct the workload. Unlike workload forensics methods implemented at the operating system (OS) and/or hypervisor level, whose data logging and monitoring mechanisms may be compromised through software attacks, this approach is implemented directly in hardware and is, therefore, immune to such attacks. The proposed method is demonstrated on an experimentation platform which consists of a 32-bit x86 architecture running Linux operating system, implemented in the Simics simulation environment. Experimental results using the Mibench workload benchmark suite reveal an overall workload identification accuracy of 96.97% at an estimated logging rate of only 5.17 KB/sec.
展开▼
