Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > comp.programming.threads > #4928

About FlushProcessWriteBuffers() and IPIs..

Show all headers | View raw

Hello..


About FlushProcessWriteBuffers() and IPIs..

It seems that the implementation of the sys_membarrier on Linux 4.3 is 
too slow. Starting with kernel 4.14, there is a new flag 
MEMBARRIER_CMD_PRIVATE_EXPEDITED that enables much faster implementation 
of the syscall using IPI.

See 
https://lttng.org/blog/2018/01/15/membarrier-system-call-performance-and-userspace-rcu/ 


for some details.

And read the following about Userspace RCU, it is also using IPIs:

membarrier system call performance and the future of Userspace RCU on Linux

Read more here:

https://lttng.org/blog/2018/01/15/membarrier-system-call-performance-and-userspace-rcu/


Cache-coherency protocols do not use IPIs, and as a user-space level 
developer you do not care about IPIs at all. One is most interested in 
the cost of cache-coherency itself. However, Win32 API provides a 
function that issues IPIs to all processors (in the affinity mask of the 
current process) FlushProcessWriteBuffers(). You can use it to 
investigate the cost of IPIs.

When i do simple synthetic test on a dual core machine I've obtained 
following numbers.

420 cycles is the minimum cost of the FlushProcessWriteBuffers() 
function on issuing core.

1600 cycles is mean cost of the FlushProcessWriteBuffers() function on 
issuing core.

1300 cycles is mean cost of the FlushProcessWriteBuffers() function on 
remote core.

Note that, as far as I understand, the function issues IPI to remote 
core, then remote core acks it with another IPI, issuing core waits for 
ack IPI and then returns.

And the IPIs have indirect cost of flushing the processor pipeline.


My C++ synchronization objects library was updated,
and now i have invented and added the scalable Fast_RWLock and
the scalable Fast_RWLockX and they are better than scalable Asymmetric
RWLocks that use IPIs, and they are costless on the reader side and
they don't use IPIs on the writer side and they are starvation-free, so 
they are really powerful, and they are now working with Windows and with 
Linux, i have tested thoroughly my C++ synchronization objects library 
and i think it is much more stable and fast.

You can read about it and download it from my website here:

https://sites.google.com/site/scalable68/c-synchronization-objects-library

The source code is inside my zip files here(they are called 
Fast_RWLockX.pas and LW_Fast_RWLockX.pas):

https://sites.google.com/site/scalable68/scalable-rwlock



Thank you,
Amine Moulay Ramdane.

Back to comp.programming.threads | Previous | Next | Find similar | Unroll thread

Thread

About FlushProcessWriteBuffers() and IPIs.. Horizon68 <horizon@horizon.com> - 2019-02-09 08:24 -0800

csiph-web