15.7 Performance Tweaks

There are a number of parameters in Simics which can be tweaked which might lead to increased performance.

• Multicore Accelerator can be enabled.
• The instruction per cycle (IPC) parameter can be decreased, see command <cpu>.set-step-rate. Similar to reducing the CPU frequency, this will cause virtual time to progress more rapidly with the same amount of instructions executed.
• Devices sometimes have a timing model that can be changed by attributes.
• When real-time performance is required, the real-time module provides some means to achieve this. See command enable-real-time-mode.
• For distributed simulation, multithreading, and Multicore Accelerator, you might want to tweak the default latency or allowed virtual time window. Increasing the latency/time window will diminish the cost of synchronizing the simulation. More information is available in chapter 16.
• In a similar way, you may want to check the time quantum used to schedule the simulated processors within a Simics process or a simulation cell. Processors unrelated to each other (they do not share memory, for example) do not need to be simulated with a small time quantum.
• The frequencies of the processors can be lowered (cpu_freq_mhz). This will cause virtual time to progress more rapidly with the same amount of instructions executed.
• If page/swap activities can be monitored on the host running Simics you might want to decrease the memory-limit. See the set-image-memory-limit command. See chapter 7.2.3.
• For large configuration using many cells and/or processors, Simics will create as many execution threads as there are available host processors and schedule the execution on these. However, sometimes it is not beneficial to use the SMT threads on the host but instead reduce the amount of created threads, using only 1 SMT thread per core. In that case, use the set-thread-limit command to limit the thread count to the available number of host cores.
• The sim->max_worker_threads attribute controls the number of threads that can be used for JIT compilation. The default number of JIT threads is automatically set conservatively low, to avoid interfering with the execution threads. This attribute can be adjusted at any time. To disable parallel JIT compilation, the sim->use_jit_threads attribute can be set to false. Then the execution threads themselves run the compilation.