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

Groups > comp.programming.threads > #2512 > unrolled thread

We have to be smart...

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

Contents

  We have to be smart... Ramine <ramine@1.1> - 2014-06-14 16:56 -0700

#2512 — We have to be smart...

Hello,

We have to be smart...


In my scalability prediction i have forget to include
the following  part of my scalable LW_RWLOCK into
the serial part of the Amdahl equation, here it is:


Read here:

http://pages.videotron.com/aminer/rwlock1.html


I have forgot to include the following part of the source code on
the writer side into the serial part of the Amdahl equation, here it is:

==
for i:=0 to GetSystemThreadCount-1 do
  begin
    while (FCount1^[i].fcount1<>0)
     do
      begin
      end;
==


So notice with me carefully on the follwing part above that the writer 
have to transfer many cache-lines from the other cores and that will 
make the serial part of the Amdahl equation expensive, add to that one 
cache-line transfer of the CAS , hence  i have redone the scalability 
prediction taking into account the above part of the source code and 
using the Amdahl equation and it has given for a small critical section 
with 3% of writes: 9X scalability on 8 cores, and  5X scalability on 16 
cores and lower and lower scalability on more and more cores , so
those scalable RWLocks are not good, RCU also is not good since it is
expensive on the writer side so it will not scale well, so i think
that RCU and my scalable RWLocks are not good candidates to choose
if you need to scale above 16 cores...


Thank you,
Amine Moulay Ramdane.

[toc] | [standalone]

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

csiph-web