Re: [Qemu-devel] GSOC 2015 Project Proposal

[Alexander Graf <agraf@suse.de>]

On 03/18/2015 12:46 PM, Stefan Hajnoczi wrote:
> On Mon, Feb 9, 2015 at 12:44 AM, Xin Tong <trent.tong@gmail.com> wrote:
>> I would like to do GSOC this summer. The project i have in mind is to
>> implement a set of facilities to make implementing Hardware
>> transactional memory (HTM) easier in QEMU.
>>
>> HTM has become available in many architecture supported by QEMU, e.g.
>> i386, PowerPC, etc. Currently, necessary memory tracking. conflict
>> detection and transaction rollbak/commit are not available in QEMU. As
>> a result, HTM is supported in a very rudimentary fashion in PowerPC,
>> i.e. the transaction begins (tbegin in PowerPC) always trigger a fault
>> to the fallback code path. Even though HTM is supported by different
>> architectures, the underlying principle are very similar and therefore
>> it is beneficial to provide a set of facilities to make implementing
>> HTM easier in QEMU.
>>
>> These facilities should include.
>>
>> A modified software TLB to make memory address and value tracking simple.
>> A performant and memory efficient value/address tracking facility to
>> detect read/write conflicts for transactions.
>> A performant and memory efficient mechanism to rollback and commit
>> memory accesses.
>> A mechanism to abort transactions on the current processor as well as
>> remote processor.
>>
>> I will come up with a more detailed proposal as application time draws
>> close. Any suggestions are appreciated at the moment.
> Hi Xin,
> Thanks for proposing this project idea.  There haven't been any
> responses yet.  I have CCed Alexander Graf and David Gibson, who have
> worked on the PowerPC target, and Richard Henderson in case he's
> interested in transactional memory.
>
> You need to find a mentor willing to supervise this project idea.
> Hopefully bumping this email thread will remind people to consider
> your idea.

I think this would be a very interesting research project, but I'm wary 
that the overhead of TLB allocation synchronization with other vcpus 
ends up pretty expensive for the normal use case of non-transactional 
operation.

I'm also not sure whether GSoC is long enough to actually get to the 
point where what you're proposing works. How much experience do you have 
in this field?


Alex