Heterogeneous System Support

From gem5
Jump to: navigation, search

(from Gabe)

I was thinking about how to set up m5 to support other architectures, and decision which has a major impact on that is whether we want to support heterogeneous systems. Here are the arguements I saw for and against.


Adding support for heterogeneous systems will be a big change. There are many systems of m5 that aren't ready to support something that's not alpha, let alone two things. If there are heterogeneous systems, the code which implements each must be present. That means that a system has to be in place to generate the appropriate pieces, which could be using templates, namespaces, or setting up standard interfaces and using a building block style approach. In all cases, this would mean carefully and totally cutting the pieces apart from one another, since it can't be assumed that any one thing will be there.


Allowing heterogeneous systems will allow simulation of mixed systems, such as a satellite embedded device talking to a server implemented on, for instance, MIPS and SPARC architectures respectively. Another setup which also becomes possible are asymmetric multiprocessing systems which use different ISAs for the different computing elements. I believe the Cell processor is an example of this. Having a heterogeneous architecture won't be as hard to implement once there have been sufficient changes to allow non-alpha architectures at all. I would estimate that the majority of the work involved will be to remove the alphacentric nature of existing code. Code set up in this way should be more modular and non-specialized, which would lead to better implementation overall.

My thoughts

I favor implementing a system which supports hetergeneous architectures, since that will be almost the same thing as supporting other architectures at all. I think being able to support multi ISA asymmetric multiprocessor systems would be a useful feature. Also, modularizing the code fully would make m5 more usable for us, and more approachable for people who want to use it for other things.

I thought this was a settled issue: as we fix all the things that need to be fixed to enable multiple ISAs with a single ISA selected at compile time, we will:

  1. not consciously do anything that makes it unreasonably difficult to add support for heterogeneous systems, and
  2. make specific features support heterogeneous systems if the amount of work is equivalent to or only slightly more than not adding that support, but
  3. not commit to supporting heterogeneous systems fully at this time since we don't need it. That is, if there are features that don't fall under rule #2, (they're needed solely for heterogeneous systems and/or are significantly more effort), then we're not going to do them right now because we have more important things to work on.

But if there aren't any necessary features that fall under rule #3, then it's possible that heterogeneous system support will just fall out.


The problem is that there is a possible contradiction between the first and second points. There very well may be situations where we'll need to go out of our way to make heterogeneous systems not require a lot of effort later. For example, the ISA compilation strategy you outlined won't work for hetergenous systems since the switch header will point to only one ISA. A system built around that will be hard to change later. The direction things are headed seems to be to not worry about heterogeneous systems, which is ok by me, but we should resolve which of those two points is dominant.


Anything that's ISA-specific will be encapsulated as a specific type that knows what ISA it belongs to. If we want to get fancier we can have the Python figure it out.

For example, FreeBSDSystem should have an AlphaFreeBSDSystem subclass to add on the Alpha-specific parts (via abstract virtual functions etc.).

-- 16:03, 7 February 2006 (EST)