<div dir="ltr"><div class="gmail_extra">On Tue, Jul 16, 2013 at 1:18 PM, Xinliang David Li <span dir="ltr"><<a href="mailto:xinliangli@gmail.com" target="_blank" class="cremed">xinliangli@gmail.com</a>></span> wrote:<br>
<div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div id=":jid" style="overflow:hidden">Ignoring FE time which can be fully parallelized and assuming 10%<br>
compile time is spent in serial module passes, 25% time is spent in<br>
CGSCC pass, the maximum speed up that can be gained by using function<br>
level parallelism is less than 3x. Even adding support for parallel<br>
compilation for leaves of CG in CGSCC pass won't help too much -- the<br>
percentage of leaf functions is < 30% in large apps I have seen.<br></div></blockquote><div><br></div><div>Can you clarify what you're basing these assumption on or how you derived your data?</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div id=":jid" style="overflow:hidden">
Module based parallelism proposed by Shuxin has max speed up of 10x,<br>
assuming body cloning does not add a lot overhead and build farm with<br>
hundred/thousands of nodes is used.</div></blockquote><div><br></div><div>Body cloning does add some overhead, so that actually needs to be measured. Also, many don't have such a build farm. <br></div></div></div></div>