<html xmlns:v="urn:schemas-microsoft-com:vml" xmlns:o="urn:schemas-microsoft-com:office:office" xmlns:w="urn:schemas-microsoft-com:office:word" xmlns:m="http://schemas.microsoft.com/office/2004/12/omml" xmlns="http://www.w3.org/TR/REC-html40">
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
<meta name="Generator" content="Microsoft Word 15 (filtered medium)">
<style><!--
/* Font Definitions */
@font-face
{font-family:Calibri;
panose-1:2 15 5 2 2 2 4 3 2 4;}
/* Style Definitions */
p.MsoNormal, li.MsoNormal, div.MsoNormal
{margin:0in;
margin-bottom:.0001pt;
font-size:12.0pt;
font-family:"Times New Roman",serif;}
a:link, span.MsoHyperlink
{mso-style-priority:99;
color:#0563C1;
text-decoration:underline;}
a:visited, span.MsoHyperlinkFollowed
{mso-style-priority:99;
color:#954F72;
text-decoration:underline;}
p
{mso-style-priority:99;
mso-margin-top-alt:auto;
margin-right:0in;
mso-margin-bottom-alt:auto;
margin-left:0in;
font-size:12.0pt;
font-family:"Times New Roman",serif;}
p.msonormal0, li.msonormal0, div.msonormal0
{mso-style-name:msonormal;
mso-margin-top-alt:auto;
margin-right:0in;
mso-margin-bottom-alt:auto;
margin-left:0in;
font-size:12.0pt;
font-family:"Times New Roman",serif;}
span.EmailStyle19
{mso-style-type:personal-reply;
font-family:"Calibri",sans-serif;
color:#1F497D;}
.MsoChpDefault
{mso-style-type:export-only;
font-family:"Calibri",sans-serif;}
@page WordSection1
{size:8.5in 11.0in;
margin:1.0in 1.0in 1.0in 1.0in;}
div.WordSection1
{page:WordSection1;}
--></style><!--[if gte mso 9]><xml>
<o:shapedefaults v:ext="edit" spidmax="1026" />
</xml><![endif]--><!--[if gte mso 9]><xml>
<o:shapelayout v:ext="edit">
<o:idmap v:ext="edit" data="1" />
</o:shapelayout></xml><![endif]-->
</head>
<body lang="EN-US" link="#0563C1" vlink="#954F72">
<div class="WordSection1">
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">I’ve done similar rankings with profile data and found it worked out pretty well.<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">In my case it was ranking function hotness but the idea was similar: sort by weight, compute various percentile cutoff points, and use those to classify. I put
in some compensation for truly degenerate cases. Consider one super-hot function that is 100x the weight of all remaining functions – I’d just call that function hot and subtract it out of the distribution<a name="_MailEndCompose"> and so spread a bit of the
hotness around elsewhere. <o:p></o:p></a></span></p>
<p class="MsoNormal"><span style="mso-bookmark:_MailEndCompose"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></span></p>
<p class="MsoNormal"><span style="mso-bookmark:_MailEndCompose"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D">Also, you should think about how to handle ties, so your sorting & ranking is “stable” somehow.<o:p></o:p></span></span></p>
<p class="MsoNormal"><span style="mso-bookmark:_MailEndCompose"><span style="font-size:11.0pt;font-family:"Calibri",sans-serif;color:#1F497D"><o:p> </o:p></span></span></p>
<span style="mso-bookmark:_MailEndCompose"></span>
<p class="MsoNormal"><b><span style="font-size:11.0pt;font-family:"Calibri",sans-serif">From:</span></b><span style="font-size:11.0pt;font-family:"Calibri",sans-serif"> llvm-dev [mailto:llvm-dev-bounces@lists.llvm.org]
<b>On Behalf Of </b>Easwaran Raman via llvm-dev<br>
<b>Sent:</b> Wednesday, December 16, 2015 2:43 PM<br>
<b>To:</b> llvm-dev@lists.llvm.org<br>
<b>Cc:</b> David Li <davidxl@google.com><br>
<b>Subject:</b> [llvm-dev] RFC: Hotness thresholds in profile header<o:p></o:p></span></p>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<p style="margin:0in;margin-bottom:.0001pt"><span style="font-size:11.0pt;font-family:"Arial",sans-serif;color:black">Hi,<br>
<br>
The problem we're trying to address:<br>
<br>
PGO transforms that are based on block counts need some way to answer isHotBlock() query. A simple comparison of the form block_count > N, for some constant N is not useful since even for the same program different profile inputs can cause different answers
for isHotBlock() on the same block. This can be addressed by comparing against a reference value that depends on the program and its input. For instance, the indexed profile header contains MaxFunctionCount and isHotBlock() could check if the block count exceeds
a certain fraction of MaxFunctionCount.</span><o:p></o:p></p>
<p style="margin:0in;margin-bottom:.0001pt"><o:p> </o:p></p>
<p style="margin:0in;margin-bottom:.0001pt"><span style="font-size:11.0pt;font-family:"Arial",sans-serif;color:black">The drawback of this approach is that it is difficult to come up with a single fraction that works both for programs with a few hot-spots and
programs with a long tail of block execution counts. Consider the following two profiles both with a MaxFunctionCount of 1000. In the first case, there are many blocks with count 100 and those (along with the function with count 1000) account for 99% of execution
time. In the second case, there are many blocks with count 1000 that together account for 99% of execution time, but there are also a non-trivial number of blocks with count 100. In the first case, we would like the hotness threshold to be 100 (or 10% of MaxFunctionCount),
but in the second case, we want it to be 1000 (100% of MaxFunctionCount). </span>
<o:p></o:p></p>
<p style="mso-margin-top-alt:0in;margin-right:0in;margin-bottom:12.0pt;margin-left:0in">
<span style="font-size:11.0pt;font-family:"Arial",sans-serif;color:black"><br>
Our proposal:<br>
<br>
We want to define hot blocks as those that account for a large percentage(P) of execution time of the program. We then make a simplistic assumption that all blocks take the same amount of time to execute. With this assumption, the execution time of the program
can be approximated by the sum of execution counts of the blocks in the program. We can then compute a maximum threshold T(P) such that execution counts of blocks whose count is >= T(P) account for P percentage of total block count.<br>
<br>
To illustrate, let {B1...B5} be the set of blocks whose counts are {1, 0, 2, 2, 5}. For this example:<br>
<br>
T(50) = 5 since {B5} accounts for 50% of total block count.<br>
T(90) = 2 since {B3, B4, B5} has a cumulative count of 9 which is 90% of total block count.<br>
T(95) = 1 since all blocks with non-zero block counts are needed to account for >=95% of total block count.<br>
<br>
We propose to record a vector of triples {P, T(P), B(P)}, where B(P) is the number of blocks with count >= T(P), for a set of pre-defined number of values of P into the header of the indexed profile. A possible set of values of P are 90, 95, 99, 99.9, 99.99,
but this should be configurable through profdata tool. The B(P) above could be useful to optimization passes in making size/performance tradeoffs. For example, if S is the average block size and S*B(95) fits well within cache, but S*B(99) well exceeds the
cache size, an optimization like loop unroller could choose to use T(95) as the threshold to avoid heavy cache misses.<br>
<br>
The {P, T(P), B(P)} triples have to be attached to the module. A common API would be added to retrieve the set of available Ps and {T(P), B(P)} pair for a given P for all flavors (front-end, IR level and sample) of profiles.
</span><o:p></o:p></p>
<p style="margin:0in;margin-bottom:.0001pt"><span style="font-size:11.0pt;font-family:"Arial",sans-serif;color:black">Thoughts?</span><o:p></o:p></p>
<p style="margin:0in;margin-bottom:.0001pt"><o:p> </o:p></p>
<p style="margin:0in;margin-bottom:.0001pt"><span style="font-size:11.0pt;font-family:"Arial",sans-serif;color:black">Thanks,</span><o:p></o:p></p>
<p style="margin:0in;margin-bottom:.0001pt"><span style="font-size:11.0pt;font-family:"Arial",sans-serif;color:black">Easwaran (with inputs from Teresa Johnson and David Li)</span><o:p></o:p></p>
</div>
</div>
</body>
</html>