[m5-users] What is a reasonable size for memory under ALPHA_FS?
Hi, I noticed that the default memory size set in Benchmarks.py is 128MB, isn't it too small for reasonable simulations? Previously when I was using ALPHA_SE, the physmem is set to 2GB, and the simulation ran well. In FS mode, however, if 2GB is used, booting up Linux (with atomic CPU) becomes extremely slow; if 1GB or 512MB is used, I can boot up the OS, start the program and make a checkpoint successfully. However, restoring from the checkpoint directly with detailed CPU (--detailed) gives me segmentation fault, the interesting thing is: if I restore the checkpoint with atomic CPU and then switch to timing and detailed ones (--standard-switch), the simulation runs well. For the default value 128MB, both --detailed and --standard-switch can run. I am confused by this observation. Am I missing anything here? What is a reasonable memory size in FS mode (say, for PARSEC programs)? Thanks, Lide ___ m5-users mailing list m5-users@m5sim.org http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
Re: [m5-users] What is a reasonable size for memory under ALPHA_FS?
I can't really answer your question about the segfault, but I think I know why ALPHA_FS and ALPHA_SE behave differently when using lots of memory. In ALPHA_SE, the benchmark will use memory as it needs it, and if there's a lot of extra memory it'll just sit there doing nothing and not affect things. In ALPHA_FS, though, Linux will clear out all the physical memory as part of the boot process. By increasing the memory size by a factor of 20, you're making that part of boot which is normally a decent part of the time take at least 20 times as long. Gabe Quoting Lide Duan leaderd...@gmail.com: Hi, I noticed that the default memory size set in Benchmarks.py is 128MB, isn't it too small for reasonable simulations? Previously when I was using ALPHA_SE, the physmem is set to 2GB, and the simulation ran well. In FS mode, however, if 2GB is used, booting up Linux (with atomic CPU) becomes extremely slow; if 1GB or 512MB is used, I can boot up the OS, start the program and make a checkpoint successfully. However, restoring from the checkpoint directly with detailed CPU (--detailed) gives me segmentation fault, the interesting thing is: if I restore the checkpoint with atomic CPU and then switch to timing and detailed ones (--standard-switch), the simulation runs well. For the default value 128MB, both --detailed and --standard-switch can run. I am confused by this observation. Am I missing anything here? What is a reasonable memory size in FS mode (say, for PARSEC programs)? Thanks, Lide ___ m5-users mailing list m5-users@m5sim.org http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
Re: [m5-users] What is a reasonable size for memory under ALPHA_FS?
There are bugs in the O3 model such that restoring directly from a checkpoint into O3 doesn't work. That's why the standard-switch model exists. I don't think it has anything to do with the memory size. Steve On Wed, Oct 13, 2010 at 11:54 AM, Lide Duan leaderd...@gmail.com wrote: Hi, I noticed that the default memory size set in Benchmarks.py is 128MB, isn't it too small for reasonable simulations? Previously when I was using ALPHA_SE, the physmem is set to 2GB, and the simulation ran well. In FS mode, however, if 2GB is used, booting up Linux (with atomic CPU) becomes extremely slow; if 1GB or 512MB is used, I can boot up the OS, start the program and make a checkpoint successfully. However, restoring from the checkpoint directly with detailed CPU (--detailed) gives me segmentation fault, the interesting thing is: if I restore the checkpoint with atomic CPU and then switch to timing and detailed ones (--standard-switch), the simulation runs well. For the default value 128MB, both --detailed and --standard-switch can run. I am confused by this observation. Am I missing anything here? What is a reasonable memory size in FS mode (say, for PARSEC programs)? Thanks, Lide ___ m5-users mailing list m5-users@m5sim.org http://m5sim.org/cgi-bin/mailman/listinfo/m5-users ___ m5-users mailing list m5-users@m5sim.org http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
Re: [m5-users] What is a reasonable size for memory under ALPHA_FS?
Additionally, in FS mode if you're running benchmarks with very large memory footprints, you can mount a swap space disk to make the system more realistic (as opposed to exorbitant amounts of simulated memory). Before running the benchmark, just run: % /sbin/swapon /dev/hdc Joel On Wed, Oct 13, 2010 at 3:20 PM, Steve Reinhardt ste...@gmail.com wrote: There are bugs in the O3 model such that restoring directly from a checkpoint into O3 doesn't work. That's why the standard-switch model exists. I don't think it has anything to do with the memory size. Steve On Wed, Oct 13, 2010 at 11:54 AM, Lide Duan leaderd...@gmail.com wrote: Hi, I noticed that the default memory size set in Benchmarks.py is 128MB, isn't it too small for reasonable simulations? Previously when I was using ALPHA_SE, the physmem is set to 2GB, and the simulation ran well. In FS mode, however, if 2GB is used, booting up Linux (with atomic CPU) becomes extremely slow; if 1GB or 512MB is used, I can boot up the OS, start the program and make a checkpoint successfully. However, restoring from the checkpoint directly with detailed CPU (--detailed) gives me segmentation fault, the interesting thing is: if I restore the checkpoint with atomic CPU and then switch to timing and detailed ones (--standard-switch), the simulation runs well. For the default value 128MB, both --detailed and --standard-switch can run. I am confused by this observation. Am I missing anything here? What is a reasonable memory size in FS mode (say, for PARSEC programs)? Thanks, Lide ___ m5-users mailing list m5-users@m5sim.org http://m5sim.org/cgi-bin/mailman/listinfo/m5-users ___ m5-users mailing list m5-users@m5sim.org http://m5sim.org/cgi-bin/mailman/listinfo/m5-users -- Joel Hestness PhD Student, Computer Architecture Dept. of Computer Science, University of Texas - Austin http://www.cs.utexas.edu/~hestness ___ m5-users mailing list m5-users@m5sim.org http://m5sim.org/cgi-bin/mailman/listinfo/m5-users