Say what TPC-C costs when SQLite commits in groups too, and with four warehouses (106)

SQLite tuned first, by the criteria committed before (and, beyond them,
checkpoints each 40 MB of WAL, in its favour). On the SD card Pure stays
4.2-10.8x ahead at equal terminals, SQLite writing 8-18x the bytes a
transaction; in RAM 3.2-3.5x. All tables equal at W=1 and W=4, and
every run consistent. In RAM the one owner is the limit (0.94 of a core
at W=4, no gain from 16 to 64 terminals), a third of its busy time
copying nodes the snapshots share; checkpoints and memory are within
their criteria.

1 files changed, 45 insertions, 0 deletions

diff --git a/research/experiments/022-tpcc-grouped/runs.sh b/research/experiments/022-tpcc-grouped/runs.sh
new file mode 100644
index 0000000..c441b36
--- /dev/null
+++ b/research/experiments/022-tpcc-grouped/runs.sh
@@ -0,0 +1,45 @@
+#!/bin/sh
+# The runs of docs/research/106, as they were made, from the repository's root
+# (the SD card's directory is any on ext4 there). Each result file records
+# SQLite's pragmas and page size in its lines.
+set -e
+out=research/experiments/022-tpcc-grouped
+sd=${SD:-$HOME/tpcc-sd}
+run() {
+    name=$1
+    shift
+    python3 benchmarks/tpcc/run.py "$@" > $out/2026-09-30-$name-linux-aarch64.jsonl
+}
+tuned='PRAGMA mmap_size=1073741824; PRAGMA cache_size=-65536; PRAGMA temp_store=MEMORY;'
+
+# SQLite's settings, W=1, SQLite alone (harness of 3c04860).
+run tune-ram-defaults --warehouses 1 --check 0 --terminals '' --sqlite-threads 1,4 --sqlite-group 16,64 \
+    --seconds 20 --sqlite-pragmas ''
+run tune-ram-tuned --warehouses 1 --check 0 --terminals '' --sqlite-threads 1,4 --sqlite-group 16,64 \
+    --seconds 20 --sqlite-pragmas "$tuned"
+run tune-sd-page4096 --directory $sd --warehouses 1 --check 0 --terminals '' --sqlite-threads 1 --sqlite-group 16,64 \
+    --seconds 30 --sqlite-pragmas "$tuned"
+run tune-sd-page1024 --directory $sd --warehouses 1 --check 0 --terminals '' --sqlite-threads 1 --sqlite-group 16,64 \
+    --seconds 30 --sqlite-page-size 1024 --sqlite-pragmas "$tuned"
+# Beyond the criteria, in SQLite's favour: a checkpoint each 40 MB of WAL.
+run tune-sd-page1024-checkpoint40000 --directory $sd --warehouses 1 --check 0 --terminals '' --sqlite-threads 1 \
+    --sqlite-group 16,64 --seconds 30 --sqlite-page-size 1024 --sqlite-pragmas "$tuned PRAGMA wal_autocheckpoint=40000;"
+run tune-ram-checkpoint10000 --warehouses 1 --check 0 --terminals '' --sqlite-threads 1,4 --sqlite-group 16,64 \
+    --seconds 20 --sqlite-pragmas "$tuned PRAGMA wal_autocheckpoint=10000;"
+
+# The matrix (harness of 20b313c, SQLite as its tuned() makes it): RAM, then
+# the card in three rounds, W=1 and W=4 taking turns.
+both='--terminals 16,64 --sqlite-threads 1,4 --sqlite-group 16,64'
+run ram-w1 --warehouses 1 --check 20000 $both --seconds 20
+run ram-w4 --warehouses 4 --check 20000 $both --seconds 20
+for round in 1 2 3; do
+    for w in 1 4; do
+        run sd-w$w-$round --directory $sd --warehouses $w --check 0 $both --seconds 30 --sqlite-page-size 1024
+    done
+done
+
+# Where the owner thread is (RAM, W=1, 64 terminals): 80 samples of its stack
+# with gdb while Pure alone runs for 90 seconds.
+python3 $out/owner-samples.py 80 $out/2026-09-30-owner-stacks-linux-aarch64.txt &
+run owner-run --warehouses 1 --check 0 --terminals 64 --sqlite-threads '' --sqlite-group '' --seconds 90
+wait