WeeFence: Toward Making Fences Free in TSO

[computer45]

Hello,

Read this:


WeeFence: Toward Making Fences Free in TSO

"Today’s fences can be quite expensive. If, instead, they were
largely free, software could benefitsubstantially: programmers could
write faster fine-grained concurrent algorithms, and C++ and Java
compilers could guarantee SC at little cost.

In this paper, we have presented WFence, a fence that is very
cheap because it allows post-fence accesses to skip it. Such accesses
can typically complete and retire before the pre-fence writes
have drained from the write buffer. If an incorrect access reordering
is about to happen, the hardware stalls for a short period to avoid
it. In addition, WFence is compatible with the use of conventional
fences in the same program.

We presented the WFence design for TSO, and compared it to
a conventional fence with speculation for 8-processor simulations.
We ran parallel kernels that contain explicit fences and parallel
applications that do not. For the kernels, WFence eliminated nearly
all of the fence stall, reducing the kernels’ execution time by an
average of 11%. For the applications, a conservative compiler algorithm
placed fences in the code to guarantee SC. Then, on average,
WFences reduced the resulting fence overhead from 38% of the
applications’ execution time to 2% (in a centralized WFence design),
or from 36% to 5% (in a distributed WFence design).
Overall, the resulting cheap fence can be a good help for parallel
programming. In our future work, we plan to optimize the distributed
GRT design for the case where a WFence maps to multiple
GRT modules."


Read more here:

http://iacoma.cs.uiuc.edu/iacoma-papers/isca13_2.pdf



Thank you,
Amine Moulay Ramdane.