Citation Context

...uires a branch predictor simulation (which is much slower than a simple profile) or hardware that reports performance information for the underlying branch predictor, such as the Profile-Me mechanism =-=[3]-=-. 2 The Icache is equivalent to a 64 kB cache, since SimpleScalar has a 64-bit instruction encoding, but we only use 32 bits for each instruction, so half the space is wasted. inst. number inst. numbe...

Citation Context

...n message-marking approaches, performance parameters are recorded in the messages as they travel through the system. Examples of this approach are ProfileMe for profiling on out-of-order proces-sors =-=[10]-=- and IP traceback for tracing the packet source [45]. In Section 2.3 we argued that message-marking approaches are more suitable for unprivileged diagnosis in Internet-like environments. Our proposed ...

Citation Context

...d to reduce contention among branches within the profile buffer, and similar schemes could be used to lower the hardware cost of our Branch Behavior Buffer. The ProfileMe approach taken by Dean et al =-=[14]-=- collects detailed information about a single sampled instruction and the interaction between a pair of sampled concurrent instructions. They capture interesting events such as cache misses and branch...

Citation Context

...is still sensible. For out-of-order machines with non-blocking caches, per-line and/or per-reference information can only be accurate or useful if some sophisticated instrumentation such as ProfileMe =-=[5]-=- on the Compaq Alpha EV67 processors and successors is used. Compiler transformations introduce problems with cost attribution. Many of the things that compilers do to improve performance confound bot...

Citation Context

...f of high-level activities such as sensing, routing, or computing, for which we need an abstraction at the appropriate granularity. Earlier work has profiled energy usage at the level of instructions =-=[8]-=-, performance events [7], program counter [13], procedures [14], processes [30], and software modules [28]. In this work, we borrow the activity abstraction of a resource principal [2, 19]. An activit...

Citation Context

...ch” techniques always obtain knowledge online and must do so for any new program they optimize. Performance counters have been extensively used for performance analysis in explaining program behavior =-=[9, 3]-=-. One of the first papers to investigate how they could be used systematically to select optimizations [23] showed impressive performance gains. However, the heuristic used was manually developed over...

Citation Context

...ts-based method to static analysis. Besides iWatcher, AccMon can also use software-based instrumentation tools such as ATOM [36] or Dyninst [16], hardware watchpoints [17, 18, 35, 37], or other tools =-=[8]-=-. However, we expect that these tools would result in significant overheads. In addition, it is possible to use special hardware [38] that provides fine-grain access control to monitor memory accesses...

Citation Context

...fferent from ours. Our goal is to build simple and easy-to-understand CPI stacks, whereas the goal for the other approaches is detailed per-instruction profiling. For example, the ProfileMe framework =-=[3]-=- randomly samples individual instructions and collects cycle-level information on a per-instruction basis. Collecting aggregate CPI stacks can be done using the ProfileMe framework by profiling many r...