Path: ...!eternal-september.org!feeder3.eternal-september.org!news.eternal-september.org!.POSTED!not-for-mail
From: anton@mips.complang.tuwien.ac.at (Anton Ertl)
Newsgroups: comp.arch
Subject: transparent huge pages (was: Address bits again)
Date: Thu, 05 Sep 2024 13:45:13 GMT
Organization: Institut fuer Computersprachen, Technische Universitaet Wien
Lines: 20
Message-ID: <2024Sep5.154513@mips.complang.tuwien.ac.at>
References: <vb9r4g$2o1f$1@gal.iecc.com> <memo.20240904163805.19028V@jgd.cix.co.uk> <Cu0CO.120079$8jx2.14596@fx17.iad>
Injection-Date: Thu, 05 Sep 2024 15:55:36 +0200 (CEST)
Injection-Info: dont-email.me; posting-host="f078a70d8fd09bf20845f2d2f555f93d";
	logging-data="382148"; mail-complaints-to="abuse@eternal-september.org";	posting-account="U2FsdGVkX19KNMnLl9iRrud9yAmJmhbA"
Cancel-Lock: sha1:PNzakBAXMhbA/mFs32pOSOM6vlY=
X-newsreader: xrn 10.11
Bytes: 1919

scott@slp53.sl.home (Scott Lurndal) writes:
>Linux supports THP (Transparent Huge Pages) were the OS automatically
>updates the translation table to coalsce smaller pages into larger blocks
>on intel/amd/arm64 processors.

I have seen good effects from THP a number of years ago: I has seen a
matrix multiply program miss the TLB every time on one of its input
matrices.  And then I ran it again, and the effect was gone.
Eventually I found out that this is due to THP, and I have to disable
THP if I want to demonstrate that effect to my students.

Recently, I thought that by using a 2MB size for an mmap()-allocated
memory block I would get THP on Linux.  Unfortunately, mmap() does not
aligng 2MB blocks to 2MB boundaries, and if it does not, the block is
not eligible for THP.  This was disappointing.

- anton
-- 
'Anyone trying for "industrial quality" ISA should avoid undefined behavior.'
  Mitch Alsup, <c17fcd89-f024-40e7-a594-88a85ac10d20o@googlegroups.com>