I have observed similar behavior for 4GB and 16GB runs with PARSEC. However I am not able to explain it either. Can anyone shed some light on this behavior?
Some answers to the below questions 1. Marss does not handle page fault. It calls qemu function cpu_x86_handle_mmu_fault() through the function Context::try_handle_fault() in sim/ptl-qemu.cpp interface. I don't think Marss adds any timing delay for page faults 2. Linux page fault handler handles the allocation of a physical frame on a page fault. More info here http://wiki.osdev.org/Paging#Page_Faults marss model is quite similar to the operation in the link above except the latencies may not be added. Regards, Adarsh Patil HPC Lab Indian Institute of Science, Bangalore India ----- Original Message ----- From: "Yuhua Guo" <[email protected]> To: [email protected] Sent: Wednesday, December 2, 2015 8:35:53 AM Subject: [marss86-devel] The relation between page fault and memory size Hello, I add some statistics of page fault accord to this mail http://comments.gmane.org/gmane.comp.emulators.marss86/746 . And I run PARSEC benchmark suite twice with 2GB and 4GB memory size respectively. However, enlarging the memory size does not improve the system performance (i.e. IPC) of different benchmarks in the PARSEC. Some benchmarks are improved, the others are degraded up to 20%. I also check the statistics of page fault, some of them increase. It does not make sense. Per my understanding, enlarging memory size should not increase page fault rate and degrade the system performance. Could you help to explain it? Thank you. And I also have following questions, thank you for answering in advance. It is better if you could help to point out the source code. 1. How does Marssx86 to load a new page from disk and mimic the latency? I read the source code, the handle_common_load_store_exceptions() seems to implement this but I do not find the latency of page fault. 2. How does Marssx86 take memory size into consideration while allocating a new page? If there is no free page, some page will be swapped out and wrote back to disk. Thank you, Yuhua Guo -- This message has been scanned for viruses and dangerous content by MailScanner , and is believed to be clean. _______________________________________________ http://www.marss86.org Marss86-Devel mailing list [email protected] https://www.cs.binghamton.edu/mailman/listinfo/marss86-devel -- This message has been scanned for viruses and dangerous content by MailScanner, and is believed to be clean.
_______________________________________________ http://www.marss86.org Marss86-Devel mailing list [email protected] https://www.cs.binghamton.edu/mailman/listinfo/marss86-devel
