From d356b5a7097ed4a821e7d496cb8359d8ac2422e4 Mon Sep 17 00:00:00 2001
From: Nicolas Frisby <nick.frisby@iohk.io>
Date: Fri, 13 Sep 2024 10:21:41 -0700
Subject: [PATCH] WIP second attempt

---
 scripts/estimating-validation-bitrate/README  | 70 -------------------
 .../estimating-validation-bitrate/README.md   | 17 +++++
 .../cumuboth.awk                              | 10 +++
 scripts/estimating-validation-bitrate/gp.scr  | 16 -----
 scripts/estimating-validation-bitrate/gp.scr2 |  4 --
 scripts/estimating-validation-bitrate/plot.gp | 17 +++++
 .../stream-then-plot.sh                       | 13 ++++
 .../streaming.sh                              | 22 ++++++
 8 files changed, 79 insertions(+), 90 deletions(-)
 delete mode 100644 scripts/estimating-validation-bitrate/README
 create mode 100644 scripts/estimating-validation-bitrate/README.md
 create mode 100644 scripts/estimating-validation-bitrate/cumuboth.awk
 delete mode 100644 scripts/estimating-validation-bitrate/gp.scr
 delete mode 100644 scripts/estimating-validation-bitrate/gp.scr2
 create mode 100644 scripts/estimating-validation-bitrate/plot.gp
 create mode 100644 scripts/estimating-validation-bitrate/stream-then-plot.sh
 create mode 100644 scripts/estimating-validation-bitrate/streaming.sh

diff --git a/scripts/estimating-validation-bitrate/README b/scripts/estimating-validation-bitrate/README
deleted file mode 100644
index d46ddc15a4..0000000000
--- a/scripts/estimating-validation-bitrate/README
+++ /dev/null
@@ -1,70 +0,0 @@
-- `db-analyser --benchmark-ledger-ops` lists the fixed stats, and then for Byron and Shelley it currently also lists the number of txs and the total size of the txs.
-
-- `db-analyser --show-block-header-size` lists the size of the block's header.
-  I ran a patched version that also simply includes the block size.
-
-- `MsgBlock` has an overhead of 2 bytes (ie the list length and the word tag, since both are <24).
-
-- I _think_ `network-mux`'s SDU overhead is 8 bytes, from https://github.com/IntersectMBO/ouroboros-network/blob/db61131c2f375842f0930a8a9cf7f83b0cb80992/network-mux/src/Network/Mux/Codec.hs#L28-L40.
-  However, I don't know how many SDUs each byte requires.
-  So I'll omit this.
- 
-Thus the number of bytes in each block that are carried by the `network-mux` is: 2 + 1 + hdrSize + txSize.
-
------
-
-The basic idea of this simple script is to divide real-world time up into non-overlapping 10 second chunks.
-We use only mutator time, essentially assuming that GC is instantaneous; this conservatively _over_-estimates the effective bit rate.
-Map each chunk to the set of blocks whose validation began during that chunk.
-Then divide the sum of the validation mutator duration for each block by the sum of the on-the-wire size of each.
-
-First, a sanity check.
-
-```
-$ cat show-block-header-size.txt | awk '($4 != "SlotNo") {print $0} ($6 != "header") {print $0}'
-[0.834136s] Started ShowBlockHeaderSize
-[0.834136s] Started ShowBlockHeaderSize
-[1926.666379s] Maximum encountered header size = 1012
-[1926.666379s] Maximum encountered header size = 1012
-[1926.666504s] Done
-[1926.666504s] Done
-ImmutableDB tip: At (Block {blockPointSlot = SlotNo 133220620, blockPointHash = 8b0597e7cf5c65b9d00af8a162d30eaae6647f868224004b02d7bd30e1d3f93f})
-ImmutableDB tip: At (Block {blockPointSlot = SlotNo 133220620, blockPointHash = 8b0597e7cf5c65b9d00af8a162d30eaae6647f868224004b02d7bd30e1d3f93f})
-```
- 
-Then, the real plot as well as another sanity check of the number of blocks mapped to each 10 second chunk of mutator time.
-
-The assumed inputs are benchmark-ledger-ops.csv and patched-show-block-header-sizes.txt, where the latter used a patched `db-analyser` to include an additional column for `GetBlockSize`.
-(The sum of the `txSizes` in the `..era-specific stats` is a slight underestimate.)
-
-```
-$ cat benchmark-ledger-ops.csv | cut -d'   ' -f1,12,13 | tail -n+2 >SlotNo-BodyTickDur-BodyAppDur.txt
-$ paste SlotNo-BodyTickDur-BodyAppDur.txt <(tail -n+2 patched-show-block-header-sizes.txt)  | gawk 'BEGIN {CONVFMT="%.18g" } ($1 != $8) { exit} {x = x + $2 + $3; y = y + 2 + $14; print $1, x, y}' > SlotNo-CumuDur-CumuSize.txt
-$ cat SlotNo-CumuDur-CumuSize.txt | gawk 'BEGIN {w = 10; CONVFMT="%.18g"; prevX = -1; prevY = 0; prevZ = 0} {x = int($2 / (1000 * 1000 * w)); y = $3} (x != prevX) {print $1, (y - prevY) / w * 8 / (1000*1000), NR - prevZ; prevY = y; prevZ = NR} {prevX = x}' >catch
-$ tail -n1 SlotNo-CumuDur-CumuSize.txt 
-133075481 140291882237 190610147733
-$ head -n1000 gp.scr*
-==> gp.scr <==
-set multiplot layout 2,1
-
-set title  'Validation Bit Rate (on an AMD EPYC 7702P)'
-
-set xlabel 'slot number of the last block in each 10 second chunk (millions of slots)'
-set ylabel 'bytes validated per 10 second chunk of mutator time (megabits per second)'
-
-plot 'catch' using 1:2 notitle
-
-set yrange [0:25]
-
-unset title
-set xlabel 'same'
-set ylabel 'same, clamped to 25'
-
-plot 'catch' using 1:($2 > 25 ? 25 : $2) notitle
-
-==> gp.scr2 <==
-set xlabel 'slot number of the last block in each 10 second chunk (millions of slots)'
-set ylabel 'blocks per 10 second chunk of mutator time, clamped to 1000'
-
-plot 'catch' using 1:(1000 < $3 ? 1000 : $3) notitle
-```
diff --git a/scripts/estimating-validation-bitrate/README.md b/scripts/estimating-validation-bitrate/README.md
new file mode 100644
index 0000000000..bea529f20c
--- /dev/null
+++ b/scripts/estimating-validation-bitrate/README.md
@@ -0,0 +1,17 @@
+- `db-analyser --benchmark-ledger-ops` lists the fixed stats, including validation times and the full blockSize.
+
+- `MsgBlock` has an overhead of 2 bytes (ie the list length and the word tag, since both are <24).
+
+- I _think_ `network-mux`'s SDU overhead is 8 bytes, from https://github.com/IntersectMBO/ouroboros-network/blob/db61131c2f375842f0930a8a9cf7f83b0cb80992/network-mux/src/Network/Mux/Codec.hs#L28-L40.
+  However, I don't know how many SDUs each byte requires.
+  So I'll omit this.
+ 
+Thus the number of bytes on-the-wire is sufficiently dominated by blockSize.
+
+-----
+
+The `nix-shell -p gawk gnuplot --run 'source stream-then-plot.sh'` command renders images that help roughly answer the question: will a full buffer containing B blocks be able to refill before its entire contents is validated?
+(As of slot 134028831, it runs for about about 3.5 minutes on my laptop.)
+
+The image width is as great as Cairo would allow.
+If you open them in Firefox and then left-click, it will zoom to full height; then you can scroll along the x-axis.
diff --git a/scripts/estimating-validation-bitrate/cumuboth.awk b/scripts/estimating-validation-bitrate/cumuboth.awk
new file mode 100644
index 0000000000..0260a178c5
--- /dev/null
+++ b/scripts/estimating-validation-bitrate/cumuboth.awk
@@ -0,0 +1,10 @@
+BEGIN { CONVFMT = "%.18g"; }
+
+{
+  SlotNo = $1; Microseconds = $12 + $13; Bytes = $14;
+
+  CumuBytes        = CumuBytes        + Bytes       ;
+  CumuMicroseconds = CumuMicroseconds + Microseconds;
+
+  print SlotNo, CumuMicroseconds, CumuBytes;
+}
diff --git a/scripts/estimating-validation-bitrate/gp.scr b/scripts/estimating-validation-bitrate/gp.scr
deleted file mode 100644
index 617f9f8314..0000000000
--- a/scripts/estimating-validation-bitrate/gp.scr
+++ /dev/null
@@ -1,16 +0,0 @@
-set multiplot layout 2,1
-
-set title  'Validation Bit Rate'
-
-set xlabel 'slot number of the last block in each 10 second chunk (millions of slots)'
-set ylabel 'bytes validated per 10 second chunk of mutator time (megabits per second)'
-
-plot 'catch' using ($1 / 1000000):2 notitle
-
-set yrange [0:25]
-
-unset title
-set xlabel 'same'
-set ylabel 'same, clamped to 25'
-
-plot 'catch' using ($1 / 1000000):($2 > 25 ? 25 : $2) notitle
diff --git a/scripts/estimating-validation-bitrate/gp.scr2 b/scripts/estimating-validation-bitrate/gp.scr2
deleted file mode 100644
index 96db6f345a..0000000000
--- a/scripts/estimating-validation-bitrate/gp.scr2
+++ /dev/null
@@ -1,4 +0,0 @@
-set xlabel 'slot number (at end of each 10 second chunk)'
-set ylabel 'blocks per 10 second chunk of mutator time, clamped to 1000'
-
-plot 'catch' using 1:(1000 < $3 ? 1000 : $3) notitle
diff --git a/scripts/estimating-validation-bitrate/plot.gp b/scripts/estimating-validation-bitrate/plot.gp
new file mode 100644
index 0000000000..d3f940dc60
--- /dev/null
+++ b/scripts/estimating-validation-bitrate/plot.gp
@@ -0,0 +1,17 @@
+if (!exists('prefix')) prefix = 'catch'
+if (!exists('suffix')) suffix = ''
+
+set terminal pngcairo transparent enhanced size 32767, 1024
+set output 'plot'.suffix.'.png'
+
+set grid ytics
+set xtics 500
+
+set xlabel 'total duration of validation (s)'
+set ylabel 'megabits per second'
+
+sizes = '10 100 1000'
+
+# FYI: words() gives length and word(,) extracts one word
+
+plot for [i=1:words(sizes)] prefix.'-'.word(sizes, i) using ($1/1000000):($2*8 < 100 ? $2*8 : 100) title word(sizes, i).' block buffer'
diff --git a/scripts/estimating-validation-bitrate/stream-then-plot.sh b/scripts/estimating-validation-bitrate/stream-then-plot.sh
new file mode 100644
index 0000000000..65e8079963
--- /dev/null
+++ b/scripts/estimating-validation-bitrate/stream-then-plot.sh
@@ -0,0 +1,13 @@
+# Crunch all the data.
+for i in 10 100 1000; do B=$i . streaming.sh & done; wait
+
+# Split the x-axis in half, so the plots are more legible.
+#
+# 125000 seconds is _currently_ the end of the x-axis if I plot the whole data set in one image.
+for i in 10 100 1000; do cat catch-$i | awk '($1/1000000 <  125000/2) {print $0}' > catch1-$i & done
+for i in 10 100 1000; do cat catch-$i | awk '($1/1000000 >= 125000/2) {print $0}' > catch2-$i & done
+wait
+
+# Render plot-1.png and plot-2.png.
+for i in 1 2; do gnuplot -e "prefix='catch$i'" -e "suffix='-$i'" plot.gp & done
+wait
diff --git a/scripts/estimating-validation-bitrate/streaming.sh b/scripts/estimating-validation-bitrate/streaming.sh
new file mode 100644
index 0000000000..2cc8a8b92e
--- /dev/null
+++ b/scripts/estimating-validation-bitrate/streaming.sh
@@ -0,0 +1,22 @@
+### Will a full buffer of size B refill before it empties?
+
+out=catch-$B
+
+# The resulting file has 3 columns: SlotNo, CumuMicroseconds, CumuBytes.
+tail -n+2 benchmark-ledger-ops.csv | awk -f cumuboth.awk >$out
+
+# Discard Byron.
+# cat $out | awk '($1 >= 4492800) { print $0 }' >$out.tmp; mv $out.tmp $out
+
+# Time and space sizes of windows of B-blocks
+#
+# ChainSel and BlockFetch clients use a buffer of 10 blocks. On top of that,
+# BlockFetch itself is buffered according to the low/high watermark, which are
+# at least 192 kibibytes and 384 kibibytes, respectively. This logic here only
+# considers the block-counted buffer, not the bytes in-flight.
+paste $out <(tail -n+$((B + 1)) $out) | awk '(NF > 3) {print $2, $5 - $2, $6 - $3}' >$out.tmp; mv $out.tmp $out
+
+# The scatter plot of this data informs the question: assuming the buffer is
+# currently full, what bit rate would be necessary in order to completely
+# refill the buffer before it empties.
+paste $out <(tail -n+2 $out) | awk '(NF > 3) {print ($1 + $4) / 2, $6 / $2}' >$out.tmp; mv $out.tmp $out