Difference between revisions of "Nate's Wish List"
From gem5
(→Memory System) |
Cirosantilli (talk | contribs) |
||
(2 intermediate revisions by one other user not shown) | |||
Line 11: | Line 11: | ||
* Handle reading and writing of gzipped files | * Handle reading and writing of gzipped files | ||
* Make switching between CPU models way easier | * Make switching between CPU models way easier | ||
+ | * Get simulations with > 100 cores working | ||
=== CPU Models === | === CPU Models === | ||
Line 25: | Line 26: | ||
* Clean up SimObject.py significantly | * Clean up SimObject.py significantly | ||
* Clean up configs/* and make it much more modular and library like | * Clean up configs/* and make it much more modular and library like | ||
+ | * Make a top level DictParam that is like a VectorParam | ||
=== Memory System === | === Memory System === | ||
Line 38: | Line 40: | ||
* Add host Linux file system support so the guest can easily access files on the host. | * Add host Linux file system support so the guest can easily access files on the host. | ||
** Maybe support the host side of vmhgfs since the guest side already exists and is GPL. | ** Maybe support the host side of vmhgfs since the guest side already exists and is GPL. | ||
+ | ** Now doable with 9p. | ||
=== Disk Images === | === Disk Images === |
Latest revision as of 09:29, 26 September 2018
Contents
Framework
- Parallelize M5
- Converting everything to support the new Params mechanism is a prerequisite
- Make various global structures thread safe (e.g. Instruction cache, Statistics, etc.)
- Unify full system and syscall emulation modes into a single build
- Unify all ISAs into a single build
- This requires fixes to the way endianness is handled in the memory system.
- Add a NOISA option that doesn't have CPUs till this happens.
- Modularize the build, allowing people to pick and choose which SimObjects get compiled in.
- Make Python/C++ I/O systems work much better together. i.e. C++ can cprintf to python file objects, or Python can print to C++ iostreams.
- Handle reading and writing of gzipped files
- Make switching between CPU models way easier
- Get simulations with > 100 cores working
CPU Models
- Support something resembling direct execution using KVM.
- Have a very fast model that dynamically compiles like PIN (or even use PIN).
- Make a configurable in-order model
- Make a very fast functional CPU model that doesn't have any stats or anything like that and is as fast as possible.
- Clean up wakeup/sleeping/interrupts so the CPU models are more consistent and the interface isn't so horrendous.
Monitoring
- Listen on a socket with cherrypy and allow people to browse and query the status of a simulation.
Configuration
- Clean up SimObject.py significantly
- Clean up configs/* and make it much more modular and library like
- Make a top level DictParam that is like a VectorParam
Memory System
- Clean up interrupt handling and route interrupts through the memory system as packets.
- Implement something like SLICC for M5.
- Implement the ARM instruction.
- Add support to flush caches.
Disk Models
- Support something like qemu's qcow2 disk image
- This is actually bsd code, so we can steal it
- Figure out how to mount it natively on linux so we don't have to convert back and forth to add stuff
- Add host Linux file system support so the guest can easily access files on the host.
- Maybe support the host side of vmhgfs since the guest side already exists and is GPL.
- Now doable with 9p.
Disk Images
- Just clean this stuff up
- Figure out a replacement for ptxdist.
- Currently think Gentoo is probably the best option.
Statistics
- Make all stats types actually work.
- Move lots of stats processing to python. Keep collection in C++, but all of the printing stuff should go to python.
- Add a sqlite backend.
- Clean up and better integrate the graphing stuff.
Ethernet
- Add Ethernet overheads (preamble, trailer, inter frame gap) to EtherLink model.
- Build and Ethernet switch model.
- Get EtherTap working.
- Keep the stuff that must run as root separate.