From: Richard W.M. Jones Date: Wed, 17 Apr 2024 10:49:45 +0000 (+0100) Subject: 2024-tcg: Add some performance facts X-Git-Url: http://git.annexia.org/?p=riscv-talks.git;a=commitdiff_plain;h=refs%2Fheads%2Fmaster 2024-tcg: Add some performance facts --- diff --git a/2024-tcg/03-not-obvious.png b/2024-tcg/03-not-obvious.png index 0129232..145ff40 100644 Binary files a/2024-tcg/03-not-obvious.png and b/2024-tcg/03-not-obvious.png differ diff --git a/2024-tcg/03-not-obvious.svg b/2024-tcg/03-not-obvious.svg index feceb27..4028c62 100644 --- a/2024-tcg/03-not-obvious.svg +++ b/2024-tcg/03-not-obvious.svg @@ -26,9 +26,9 @@ inkscape:pagecheckerboard="0" inkscape:deskcolor="#d1d1d1" inkscape:document-units="mm" - inkscape:zoom="1.0393815" - inkscape:cx="396.38958" - inkscape:cy="422.84763" + inkscape:zoom="0.73495371" + inkscape:cx="397.30394" + inkscape:cy="422.47559" inkscape:window-width="1920" inkscape:window-height="1029" inkscape:window-x="0" @@ -100,7 +100,7 @@ style="stroke-width:0.264583" x="21.956394" y="70.401794" - id="tspan11"> might be worse + 1 in 6 jumps are slow pathbut lookup_tb/tb_lookup takes 17% total time diff --git a/2024-tcg/04-experiments.png b/2024-tcg/04-experiments.png index a323734..04c38fa 100644 Binary files a/2024-tcg/04-experiments.png and b/2024-tcg/04-experiments.png differ diff --git a/2024-tcg/04-experiments.svg b/2024-tcg/04-experiments.svg index 2afaa97..a92c520 100644 --- a/2024-tcg/04-experiments.svg +++ b/2024-tcg/04-experiments.svg @@ -26,9 +26,9 @@ inkscape:pagecheckerboard="0" inkscape:deskcolor="#d1d1d1" inkscape:document-units="mm" - inkscape:zoom="1.0393815" - inkscape:cx="396.38958" - inkscape:cy="422.84763" + inkscape:zoom="1.4142136" + inkscape:cx="381.13056" + inkscape:cy="199.75767" inkscape:window-width="1920" inkscape:window-height="1029" inkscape:window-x="0" @@ -95,16 +95,96 @@ style="stroke-width:0.264583" x="21.956394" y="64.228172" - id="tspan20"> * CPU state optimization * TB lookup changes * CPU state optimization * QOM casts + +5% improvement + 14% of total time page table walking + jmp cache up to 64 addrs per page: - removing 47% worse - other adjustments made it worse + super-page = 4x pages => 33% slower 8x pages => 50% slower + 10% of total time diff --git a/2024-tcg/05-lessons.png b/2024-tcg/05-lessons.png index cb96155..a6102c0 100644 Binary files a/2024-tcg/05-lessons.png and b/2024-tcg/05-lessons.png differ diff --git a/2024-tcg/05-lessons.svg b/2024-tcg/05-lessons.svg index 31c2d3b..da282c4 100644 --- a/2024-tcg/05-lessons.svg +++ b/2024-tcg/05-lessons.svg @@ -118,5 +118,16 @@ x="21.956394" y="88.922668" id="tspan27"> * Flame graphs & profiles + (63% slower when 16 vCPUs pinned to 8 pCPUs)