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

Groups > comp.programming.threads > #2774 > unrolled thread

We have to be more precise in computer science...

Back to article view | Back to comp.programming.threads

Contents

  We have to be more precise in computer science... Ramine <ramine@1.1> - 2014-12-08 19:09 -0800

#2774 — We have to be more precise in computer science...

Hello,


We have to be more precise in computer science...

To be more sure i have just benchmarked a cache-line
transfer between cores and you will not believe it !
it is so expensive on x86 ! cause it takes around
800 CPU cycles , so i think i have correctly reasoned
when i have said that the following reader-writer lock
(of Joe Duffy an architect at Microsoft) is a garbage.

So please reread carefully my reasonning that follows
cause it is correct and true:


I must say that we have to be carefull, because i have just
read the following webpage about a more scalable reader/writer lock
by an architect at microsoft called Joe Duffy... but you have to be 
carefull because this reader/writer lock is not really scalable,
it is a garbage, and i will think as an architect and explain to you why...

Here is the webpage, and my explanation follows...


http://joeduffyblog.com/2009/02/20/a-more-scalable-readerwriter-lock-and-a-bit-less-harsh-consideration-of-the-idea/


So look inside the EnterWriteLock() of the reader/writer above,
you will notice that it is first executing Interlocked.Exchange(ref 
m_writer, 1), that means it is atomicaly making m_writer equal 1,
so that to block readers from entering the reader section,
but this is garbage, cause look after that he is doing this:

for (int i = 0; i < m_readers.Length; i++)

while (m_readers[i].m_taken != 0) sw.SpinOnce();


So after making m_writers equal 1 so that to block the readers,
he is transfering  many cache-lines between cores, and this is really 
expensive and it will make the serial part of the Amdahl's law bigger 
and bigger when more and more cores will be used , so this will not 
scale, so it is garbage.

The Dmitry Vyukov distributed reader-writer mutex doesn't have this
weakness, because look at the source code here:

http://www.1024cores.net/home/lock-free-algorithms/reader-writer-problem/distributed-reader-writer-mutex


Because he is doing this on the distr_rw_mutex_wrlock() side:

for (i = 0; i != mtx->proc_count; i += 1)
         pthread_rwlock_wrlock(&mtx->cell[i].mtx);


So we have to be smart here and notice with me that as the "i" counter 
variable goes from 0 to proc_count, the reader side will still be 
allowed to enter and to enter again the reader section on scenarios with 
more contention, so in contrast with the above reader-writer lock, this 
part of the distributed lock is not counted as only a serial part of the 
Amdahl's law, because it allows also the reader threads to enter
and to enter again the reader section, so this part contains a parallel 
part of the Amdahl's law, and this makes this distributed reader-writer 
lock to effectively scale. That's even better with my Distributed 
sequential lock , because it scales even better than the distributed 
reader-writer mutex of Dmitry Vyukov.



Hope you have understood my architect way of thinking.



Thank you for your time.



Amine Moulay Ramdane.

[toc] | [standalone]

Back to top | Article view | comp.programming.threads

csiph-web