Held memory is memory.current less the page cache, which the kernel hands to another tenant when the host needs it. The score charges memory.current, as spec section 8.5 says.
When
Total
Anonymous
Shared
Page cache
Kernel
idle
0.10
0.01
0.00
0.07
0.00
peak
0.10
0.01
0.00
0.07
0.00
ready
0.03
0.01
0.00
0.02
0.00
ready_repetition
0.01
0.00
0.00
0.00
0.00
Latency, warm single client
Class
Operation
p50 ms
p99 ms
Target ms
R1
point
0.12
0.25
10
R2
filter
0.22
0.45
200
R3
neighbourhood
0.15
0.40
50
R4
deep
0.51
0.72
500
R5
path
1.35
3.32
500
R6
pattern
0.22
0.47
2,000
R7
analytic
17.61
24.34
5,000
W1
insert
1.16
3.65
20
W2
insert edge
0.46
0.57
10
W3
batch
20.88
36.19
500
U1
modify
0.84
1.32
20
U2
rewire
1.16
1.52
30
D1
delete edge
0.84
1.38
20
D2
delete cascade
11.71
15.33
100
O1
person by email
1.01
1.64
20
O2
accounts by opening date
0.99
1.39
200
O3
people by surname and city
1.14
1.33
100
S1
amount band
11.12
11.90
1,000
S3
text search
4.42
4.68
2,000
J4
neighbours alike
3.19
4.82
500
R1w
point, worst case
0.14
0.21
10
R2w
filter, worst case
0.57
1.41
200
R3w
neighbourhood, worst case
0.51
1.57
50
R4w
deep, worst case
0.59
0.96
500
R5w
path, worst case
1.58
4.01
500
R6w
pattern, worst case
0.18
0.35
2,000
R7w
analytic, worst case
17.47
18.67
5,000
Operational surface: 10 points
Item
Points
Source
index dp_account_country_status_balance: CREATE INDEX "dp_account_country_status_balance" ON "account" ("country", "status", "balance")
1
audited
index dp_account_owner_id: CREATE INDEX "dp_account_owner_id" ON "account" ("owner_id")
1
audited
index dp_knows_dst_id: CREATE INDEX "dp_knows_dst_id" ON "knows" ("dst_id")
1
audited
index dp_knows_src_id: CREATE INDEX "dp_knows_src_id" ON "knows" ("src_id")
1
audited
index dp_transfer_dst_id: CREATE INDEX "dp_transfer_dst_id" ON "transfer" ("dst_id")
1
audited
index dp_transfer_src_id_ts: CREATE INDEX "dp_transfer_src_id_ts" ON "transfer" ("src_id", "ts")
# SQLite 3.53.4, the C library, served by dp-sqlite (adapters/sqlite/server).
#
# SQLite is a library with no server, so the adapter brings the smallest one
# that puts it on a network, and runs it inside the SUT's container: each TCP
# connection is one SQLite connection, so SQLite's CPU, memory and disk land
# in the container's cgroup like every other system's. The library is the
# amalgamation github.com/mattn/go-sqlite3 v1.14.52 bundles, SQLite 3.53.4
# (2026-07-24, source id bf7c7f30031888f4e796e429ab3978879485813aaca6f641c7b33e4e09459bcc),
# compiled with that package's options; dump_config reports every one as
# compile_options.
#
# SQLite publishes no image, so the adapter builds one on the engine's host:
# `adapters/sqlite/build-image.sh` (Containerfile beside this file) builds
# localhost/dpbench-sqlite and prints its digest. The image is local, so
# `image` pins the digest that build printed.
#
# THE DIGEST BELOW IS A PLACEHOLDER. Run build-image.sh on the engine's host
# and put the digest it prints here; until then P0 cannot find the image.
name: sqlite
system: SQLite
version: "3.53.4"
image: localhost/dpbench-sqlite@sha256:e07b2b03bf92461e6d41d1a9e7d95437e0ab91affeced74b8130a25c5adb1467
model: relational
protocols: [json-lines]
query_language: sql
# A commit is durable when it returns (spec §9.2). In WAL mode SQLite syncs
# the log on every commit at synchronous FULL, which is SQLite's default
# (SQLITE_DEFAULT_SYNCHRONOUS=2). This build lowers the default for WAL
# databases to NORMAL (SQLITE_DEFAULT_WAL_SYNCHRONOUS=1), where the last
# commits before a power loss can roll back, so the server keeps FULL and the
# arguments below say so; it is needed only for durability, and exempt.
# SQLite's transactions are serializable: one writer at a time, and each
# reader sees one snapshot.
durability:
settings: { synchronous: "FULL" }
isolation: serializable
# What this configuration asks of an operator. Each line costs surface points
# (spec §10.4), and the audit checks every line against the running server:
# dump_config against a container started without the arguments below.
declared_surface:
services: 1
runtime_dependencies: []
# The PostgreSQL adapter's indexes, for the same reasons: owner_id for R1,
# R3 and D2's cascade, (country, status, balance) for R2,
# transfer(src_id, ts) for R3, R4, R6, R7 and the cascade,
# transfer(dst_id) for the cascade, knows(src_id) for R3 and R5,
# knows(dst_id) for R5 and the cascade. The keys are the canonical schema's:
# an INTEGER PRIMARY KEY is the rowid each table is ordered by, and every
# reference is a foreign key that cascades on delete.
secondary_indexes:
- "account(owner_id)"
- "account(country, status, balance)"
- "transfer(src_id, ts)"
- "transfer(dst_id)"
- "knows(src_id)"
- "knows(dst_id)"
hand_sized_caches: []
non_default_settings:
# A write-ahead log, so readers never wait for the writer and the writer
# never waits for readers. The default rollback journal is as durable,
# but each commit then waits for every reader to finish, and new readers
# wait for the commit; a reasonable operator of a database with more
# than one client turns WAL on. It persists in the database file.
journal_mode: "WAL"
# SQLite allows one writer at a time. A connection that finds the write
# lock taken waits up to 5 s for it, the timeout Python's sqlite3 module
# and go-sqlite3 default to, instead of failing at once (SQLite's
# default is 0). A write still waiting then fails with SQLITE_BUSY, which
# the driver retries as a conflict; every retry and failure is counted.
busy_timeout: "5000"
# Enforce the canonical foreign keys, so D2's DELETE cascades as the
# schema declares (spec §9.2). SQLite parses foreign keys but enforces
# none unless told to; the alternative, deleting from every referencing
# table inside D2, would leave every other write unchecked.
foreign_keys: "ON"
maintenance_tasks: []
storage_areas: [data]
# Without statistics SQLite's planner guesses each table's size; for R7 it
# guesses a scan of every transfer over the top organizations' accounts.
# With them it also skip-scans the account index for O2.
manual_steps: ["ANALYZE after import"]
envelope:
size: S
memory_limit_gib: 8
plugin: sqlite
container:
port: 6543
data_path: /data
dataset_path: /dataset
# The write-ahead log beside the database. It exists while any connection
# is open, as the adapter's is during every footprint.
log_paths: [dp.sqlite-wal]
# The image's entrypoint names the harness's paths and port. Every setting
# an operator chose is here, applied by every connection when it opens.
args: ["-pragma", "busy_timeout=5000", "-pragma", "journal_mode=WAL", "-pragma", "synchronous=FULL", "-pragma", "foreign_keys=ON"]
Executed DDL
CREATE TABLE "place" (
"id" INTEGER NOT NULL PRIMARY KEY,
"name" TEXT NOT NULL,
"type" TEXT NOT NULL,
"part_of" INTEGER REFERENCES "place" ("id") ON DELETE CASCADE
);
CREATE TABLE "person" (
"id" INTEGER NOT NULL PRIMARY KEY,
"first_name" TEXT NOT NULL,
"last_name" TEXT NOT NULL,
"birth_date" TEXT NOT NULL,
"created_at" INTEGER NOT NULL,
"risk_score" REAL NOT NULL,
"email" TEXT,
"lives_in" INTEGER NOT NULL REFERENCES "place" ("id") ON DELETE CASCADE
);
CREATE TABLE "organization" (
"id" INTEGER NOT NULL PRIMARY KEY,
"name" TEXT NOT NULL,
"type" TEXT NOT NULL,
"founded" TEXT NOT NULL,
"located_in" INTEGER NOT NULL REFERENCES "place" ("id") ON DELETE CASCADE
);
CREATE TABLE "account" (
"id" INTEGER NOT NULL PRIMARY KEY,
"owner_id" INTEGER NOT NULL REFERENCES "person" ("id") ON DELETE CASCADE,
"org_id" INTEGER REFERENCES "organization" ("id") ON DELETE CASCADE,
"opened_at" INTEGER NOT NULL,
"balance" INTEGER NOT NULL,
"status" TEXT NOT NULL,
"country" TEXT NOT NULL
);
CREATE TABLE "device" (
"id" INTEGER NOT NULL PRIMARY KEY,
"kind" TEXT NOT NULL,
"os" TEXT NOT NULL,
"first_seen" INTEGER NOT NULL,
"fingerprint" TEXT NOT NULL
);
CREATE TABLE "tag" (
"id" INTEGER NOT NULL PRIMARY KEY,
"name" TEXT NOT NULL,
"class" TEXT NOT NULL
);
CREATE TABLE "post" (
"id" INTEGER NOT NULL PRIMARY KEY,
"author_id" INTEGER NOT NULL REFERENCES "person" ("id") ON DELETE CASCADE,
"tag_id" INTEGER REFERENCES "tag" ("id") ON DELETE CASCADE,
"created_at" INTEGER NOT NULL,
"length" INTEGER NOT NULL,
"language" TEXT NOT NULL,
"content" TEXT NOT NULL
);
CREATE TABLE "knows" (
"id" INTEGER NOT NULL PRIMARY KEY,
"src_id" INTEGER NOT NULL REFERENCES "person" ("id") ON DELETE CASCADE,
"dst_id" INTEGER NOT NULL REFERENCES "person" ("id") ON DELETE CASCADE,
"since" TEXT NOT NULL,
"strength" REAL NOT NULL
);
CREATE TABLE "works_at" (
"id" INTEGER NOT NULL PRIMARY KEY,
"src_id" INTEGER NOT NULL REFERENCES "person" ("id") ON DELETE CASCADE,
"dst_id" INTEGER NOT NULL REFERENCES "organization" ("id") ON DELETE CASCADE,
"from_date" TEXT NOT NULL,
"role" TEXT NOT NULL
);
CREATE TABLE "uses" (
"id" INTEGER NOT NULL PRIMARY KEY,
"src_id" INTEGER NOT NULL REFERENCES "person" ("id") ON DELETE CASCADE,
"dst_id" INTEGER NOT NULL REFERENCES "device" ("id") ON DELETE CASCADE,
"first_seen" INTEGER NOT NULL,
"last_seen" INTEGER NOT NULL
);
CREATE TABLE "transfer" (
"id" INTEGER NOT NULL PRIMARY KEY,
"src_id" INTEGER NOT NULL REFERENCES "account" ("id") ON DELETE CASCADE,
"dst_id" INTEGER NOT NULL REFERENCES "account" ("id") ON DELETE CASCADE,
"ts" INTEGER NOT NULL,
"amount" INTEGER NOT NULL,
"currency" TEXT NOT NULL,
"channel" TEXT NOT NULL
);
CREATE TABLE "interested_in" (
"id" INTEGER NOT NULL PRIMARY KEY,
"src_id" INTEGER NOT NULL REFERENCES "person" ("id") ON DELETE CASCADE,
"dst_id" INTEGER NOT NULL REFERENCES "tag" ("id") ON DELETE CASCADE,
"weight" REAL NOT NULL
);
CREATE TABLE "likes" (
"id" INTEGER NOT NULL PRIMARY KEY,
"src_id" INTEGER NOT NULL REFERENCES "person" ("id") ON DELETE CASCADE,
"dst_id" INTEGER NOT NULL REFERENCES "post" ("id") ON DELETE CASCADE,
"ts" INTEGER NOT NULL
);
INSERT INTO "place" ("id", "name", "type", "part_of") VALUES (?, ?, ?, ?) -- each row of /dataset/data/place.csv;
INSERT INTO "person" ("id", "first_name", "last_name", "birth_date", "created_at", "risk_score", "email", "lives_in") VALUES (?, ?, ?, ?, ?, ?, ?, ?) -- each row of /dataset/data/person.csv;
INSERT INTO "organization" ("id", "name", "type", "founded", "located_in") VALUES (?, ?, ?, ?, ?) -- each row of /dataset/data/organization.csv;
INSERT INTO "account" ("id", "owner_id", "org_id", "opened_at", "balance", "status", "country") VALUES (?, ?, ?, ?, ?, ?, ?) -- each row of /dataset/data/account.csv;
INSERT INTO "device" ("id", "kind", "os", "first_seen", "fingerprint") VALUES (?, ?, ?, ?, ?) -- each row of /dataset/data/device.csv;
INSERT INTO "tag" ("id", "name", "class") VALUES (?, ?, ?) -- each row of /dataset/data/tag.csv;
INSERT INTO "post" ("id", "author_id", "tag_id", "created_at", "length", "language", "content") VALUES (?, ?, ?, ?, ?, ?, ?) -- each row of /dataset/data/post.csv;
INSERT INTO "knows" ("id", "src_id", "dst_id", "since", "strength") VALUES (?, ?, ?, ?, ?) -- each row of /dataset/data/knows.csv;
INSERT INTO "works_at" ("id", "src_id", "dst_id", "from_date", "role") VALUES (?, ?, ?, ?, ?) -- each row of /dataset/data/works_at.csv;
INSERT INTO "uses" ("id", "src_id", "dst_id", "first_seen", "last_seen") VALUES (?, ?, ?, ?, ?) -- each row of /dataset/data/uses.csv;
INSERT INTO "transfer" ("id", "src_id", "dst_id", "ts", "amount", "currency", "channel") VALUES (?, ?, ?, ?, ?, ?, ?) -- each row of /dataset/data/transfer.csv;
INSERT INTO "interested_in" ("id", "src_id", "dst_id", "weight") VALUES (?, ?, ?, ?) -- each row of /dataset/data/interested_in.csv;
INSERT INTO "likes" ("id", "src_id", "dst_id", "ts") VALUES (?, ?, ?, ?) -- each row of /dataset/data/likes.csv;
CREATE INDEX "dp_account_owner_id" ON "account" ("owner_id");
CREATE INDEX "dp_account_country_status_balance" ON "account" ("country", "status", "balance");
CREATE INDEX "dp_transfer_src_id_ts" ON "transfer" ("src_id", "ts");
CREATE INDEX "dp_transfer_dst_id" ON "transfer" ("dst_id");
CREATE INDEX "dp_knows_src_id" ON "knows" ("src_id");
CREATE INDEX "dp_knows_dst_id" ON "knows" ("dst_id");
Settings that differ from a fresh container of the same image
Setting
Image default
This run
busy_timeout
0
5000
foreign_keys
0
1
journal_mode
delete
wal
5. Queries
The adapter runs these statements, with parameters bound as $n.
O2
SELECT id FROM account
WHERE opened_at >= ?1 AND opened_at < ?2 AND balance >= ?3
S1
SELECT count(*), COALESCE(sum(amount), 0) FROM transfer
WHERE ts >= ?1 AND ts < ?2 AND channel = ?3 AND amount BETWEEN ?4 AND ?5
J4
SELECT count(DISTINCT o.id)
FROM person p
JOIN person o ON o.lives_in = p.lives_in AND o.id <> p.id
WHERE p.id = ?1
AND o.id IN (SELECT src_id FROM works_at)
AND o.id IN (SELECT io.src_id FROM interested_in io
WHERE io.dst_id IN (SELECT dst_id FROM interested_in WHERE src_id = ?1))
J1
WITH RECURSIVE r(id, depth) AS (
SELECT ?1, 0
UNION
SELECT k.dst_id, r.depth + 1 FROM r JOIN knows k ON k.src_id = r.id WHERE r.depth < 4
)
SELECT count(DISTINCT id) FROM r WHERE id <> ?1
S2
SELECT id FROM person
WHERE risk_score >= ?1 AND birth_date >= ?2 AND birth_date < ?3
S4
SELECT currency, channel, count(*), sum(amount) FROM transfer
WHERE ts >= ?1 AND ts < ?2
GROUP BY currency, channel
J5
SELECT count(*)
FROM transfer t1
JOIN transfer t2 ON t2.src_id = t1.dst_id AND t2.ts > t1.ts
JOIN transfer t3 ON t3.src_id = t2.dst_id AND t3.ts > t2.ts
JOIN transfer t4 ON t4.src_id = t3.dst_id AND t4.ts > t3.ts AND t4.ts < t1.ts + 2592000
WHERE t1.src_id = ?1 AND t1.ts >= ?2 AND t1.ts < ?3
O1
SELECT p.id, p.first_name, p.last_name, a.id, a.balance, a.status
FROM person p LEFT JOIN account a ON a.owner_id = p.id
WHERE p.email = ?1
J2
SELECT region.name, count(*), sum(t.amount)
FROM knows k
JOIN account a ON a.owner_id = k.dst_id
JOIN transfer t ON t.src_id = a.id AND t.ts >= ?2 AND t.ts < ?3
JOIN account b ON b.id = t.dst_id
JOIN person o ON o.id = b.owner_id
JOIN place city ON city.id = o.lives_in
JOIN place region ON region.id = city.part_of
WHERE k.src_id = ?1
GROUP BY region.id, region.name
S3
SELECT count(*) FROM post
WHERE language = ?1 AND created_at >= ?2 AND created_at < ?3 AND instr(content, ?4) > 0
J3
SELECT count(*)
FROM knows k1
JOIN knows k2 ON k2.src_id = k1.dst_id
JOIN knows k3 ON k3.src_id = k2.dst_id AND k3.dst_id = k1.src_id
WHERE k1.src_id = ?1 AND k1.dst_id <> ?1 AND k2.dst_id <> ?1 AND k2.dst_id <> k1.dst_id
O3
SELECT count(DISTINCT p.id), count(a.id), COALESCE(sum(a.balance), 0)
FROM person p LEFT JOIN account a ON a.owner_id = p.id
WHERE p.last_name = ?1 AND p.lives_in = ?2
R1
SELECT p.id, p.first_name, p.last_name, a.id, a.balance, a.status
FROM person p LEFT JOIN account a ON a.owner_id = p.id
WHERE p.id = ?1
R2
SELECT id FROM account
WHERE country = ?1 AND status = 'open' AND balance BETWEEN ?2 AND ?3
R3
SELECT k.dst_id, COALESCE(SUM(t.amount), 0)
FROM knows k
LEFT JOIN account a ON a.owner_id = k.dst_id
LEFT JOIN transfer t ON t.src_id = a.id AND t.ts >= ?2 AND t.ts < ?3
WHERE k.src_id = ?1
GROUP BY k.dst_id
R4
WITH RECURSIVE r(id, depth) AS (
SELECT ?1, 0
UNION
SELECT t.dst_id, r.depth + 1 FROM r JOIN transfer t ON t.src_id = r.id
WHERE r.depth < 3 AND t.ts >= ?2 AND t.ts < ?3
)
SELECT DISTINCT id FROM r WHERE id <> ?1
R5
WITH RECURSIVE
f(id, d) AS (
SELECT ?1, 0
UNION
SELECT k.dst_id, f.d + 1 FROM f JOIN knows k ON k.src_id = f.id WHERE f.d < 3
),
b(id, d) AS (
SELECT ?2, 0
UNION
SELECT k.src_id, b.d + 1 FROM b JOIN knows k ON k.dst_id = b.id WHERE b.d < 3
)
SELECT min(f.d + b.d) FROM f JOIN b ON f.id = b.id HAVING count(*) > 0
R6
WITH RECURSIVE p(acc, path, len) AS (
SELECT t.dst_id, ',' || ?1 || ',' || t.dst_id || ',', 1
FROM transfer t WHERE t.src_id = ?1 AND t.ts >= ?2 AND t.ts < ?3
UNION ALL
SELECT t.dst_id, p.path || t.dst_id || ',', p.len + 1
FROM p JOIN transfer t ON t.src_id = p.acc
WHERE p.len < 5 AND p.acc <> ?1 AND t.ts >= ?2 AND t.ts < ?3
AND (t.dst_id = ?1 OR instr(p.path, ',' || t.dst_id || ',') = 0)
)
SELECT count(*) FROM p WHERE acc = ?1 AND len >= 3
R7
WITH top AS (
SELECT org_id FROM account WHERE org_id IS NOT NULL
GROUP BY org_id ORDER BY count(*) DESC, org_id LIMIT 20
)
SELECT strftime('%Y-%m', t.ts, 'unixepoch'), sa.country, da.country, SUM(t.amount)
FROM transfer t
JOIN account sa ON sa.id = t.src_id
JOIN account da ON da.id = t.dst_id
WHERE sa.org_id IN (SELECT org_id FROM top) AND t.ts >= ?1 AND t.ts < ?2
GROUP BY 1, 2, 3
W1
BEGIN IMMEDIATE;
INSERT INTO person (id, first_name, last_name, birth_date, created_at, risk_score, email, lives_in)
VALUES (?1, ?2, ?3, '1990-01-01', ?4, 0.5, NULL, 1);
INSERT INTO account (id, owner_id, org_id, opened_at, balance, status, country)
VALUES (?1, ?2, NULL, ?3, 0, 'open', 'ZZ');
INSERT INTO knows (id, src_id, dst_id, since, strength) VALUES (?1, ?2, ?3, ?4, 0.5) -- once per edge;
COMMIT
W2
INSERT INTO transfer (id, src_id, dst_id, ts, amount, currency, channel)
VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7)
W3
BEGIN IMMEDIATE;
INSERT INTO transfer (id, src_id, dst_id, ts, amount, currency, channel)
VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7) -- once per transfer, on one prepared statement;
COMMIT
U1
UPDATE account SET balance = ?2, status = ?3 WHERE id = ?1
U2
BEGIN IMMEDIATE;
DELETE FROM works_at WHERE src_id = ?1;
INSERT INTO works_at (id, src_id, dst_id, from_date, role) VALUES (?2, ?1, ?3, ?4, 'engineer');
COMMIT
D1
DELETE FROM transfer WHERE id = ?1
D2
DELETE FROM person WHERE id = ?1
6. Concurrency ladder, mixed
k
X ops/s
Q
X × Q
1 k*
742.8
1.000
742.8
4
593.0
0.476
282.3
16
813.0
0.282
229.7
64
837.6
0.209
174.7
7. Memory over the duty cycle
memory.current of the median repetition, sampled every 100 ms. Shaded bands are the busy windows; the rest is idle.
8. Validation
0 invalid operation instances.
Write validation read back 126 account, 400 batch, 200 person alive, 200 person deleted, 200 transfer alive, 200 transfer deleted, 126 works, with 0 kinds failing.
9. Deviations
dataset.scale: 300K is not a DP-Core scale factor (spec §6.6)
tiers: 15% off-index and 5% worst-case operations blended in; spec §9.3's mix is typical operations alone
profile: quick timings: ladders, duty cycle and warm-up are shorter than spec §8.5, §8.9 and §9.5 require
repetitions: 1; spec §9.5 publishes the median of 3
every container runs --privileged (DPBENCH_PRIVILEGED): the host's kernel has no BPF syscall, so the runtime cannot attach a cgroup v2 device filter; cpu, cpuset, memory and pids limits apply as before
warm-up before P5 runs at 16 clients: spec §9.5 says k*, which only P7 finds
10. Phase detail
Import (P2): 213,727 rows in 1.1 s: load 0.9 s, keys 0.0 s, indexes 0.1 s. Peak memory during load 0.04 GiB.
Repetitions (P7, P8)
n
T
k*
Q at k*
M GiB
After work GiB
1
742.8
1
1.000
0.09
0.07
Coefficient of variation of DataPower across repetitions: 0.0%.
Memory floor (P9)
Limit GiB
Q
Outcome
8
1.000
ran
4
1.000
ran
2
1.000
ran
1
1.000
ran
Cold (P3) and warm (P5) single-client latency, ms
Class
Operation
Cold p50
Cold p99
Warm p50
Warm p99
R1
point
0.56
1.49
0.12
0.25
R2
filter
0.31
0.72
0.22
0.45
R3
neighbourhood
0.69
1.70
0.15
0.40
R4
deep
1.48
6.33
0.51
0.72
R5
path
2.00
17.34
1.35
3.32
R6
pattern
0.19
0.36
0.22
0.47
R7
analytic
16.30
18.77
17.61
24.34
W1
insert
1.37
4.19
1.16
3.65
W2
insert edge
1.03
1.22
0.46
0.57
W3
batch
13.02
14.38
20.88
36.19
U1
modify
1.53
1.82
0.84
1.32
U2
rewire
1.54
2.10
1.16
1.52
D1
delete edge
0.81
0.88
0.84
1.38
D2
delete cascade
11.77
16.93
11.71
15.33
O1
person by email
0.99
2.42
1.01
1.64
O2
accounts by opening date
1.05
1.18
0.99
1.39
O3
people by surname and city
1.12
1.18
1.14
1.33
S1
amount band
5.33
5.47
11.12
11.90
S3
text search
4.51
4.66
4.42
4.68
J4
neighbours alike
3.95
4.61
3.19
4.82
R1w
point, worst case
0.25
0.29
0.14
0.21
R2w
filter, worst case
0.58
0.83
0.57
1.41
R3w
neighbourhood, worst case
0.47
1.65
0.51
1.57
R4w
deep, worst case
0.65
0.92
0.59
0.96
R5w
path, worst case
1.98
2.54
1.58
4.01
R6w
pattern, worst case
0.22
0.22
0.18
0.35
R7w
analytic, worst case
17.61
17.87
17.47
18.67
Ladder (P6): read
k
X ops/s
Errors
Driver CPU
1
883.8
0
2.5%
4
1,740.8
0
5.0%
16
1,754.8
0
4.4%
64
1,769.4
0
4.6%
Ladder (P6): insert
k
X ops/s
Errors
Driver CPU
1
529.2
0
3.1%
4
480.2
0
2.8%
16
452.0
0
2.6%
64
371.2
0
2.3%
Ladder (P6): modify-delete
k
X ops/s
Errors
Driver CPU
1
621.6
0
1.8%
4
604.8
0
1.5%
16
586.4
0
1.7%
64
549.0
0
1.5%
Ladder (P6): write
k
X ops/s
Errors
Driver CPU
1
553.8
0
2.5%
4
508.4
0
2.4%
16
489.0
0
2.2%
64
434.0
0
2.1%
11. Run timeline
The run from its first phase to its last, every second. The driver counts every operation, ramps and warm-ups included, and the probe samples the SUT's cgroup every 100 ms. Bands mark the phases; P3, the cold run, and P4, the warm-up, stand out in teal. After the duty cycle the run asks the kernel to reclaim everything the SUT holds, and the drop in memory shows what the system gives back.
ThroughputValid operations per second, ramps and warm-ups included.Latencyp99 of every operation in each second, all classes together, on a log scale.MemoryThe SUT's cgroup memory, stacked: what the processes hold, shared memory such as a buffer pool, and the page cache the kernel can reclaim. The envelope allows 8 GiB.
process (anon)shared memorypage cachekernel
CPUCores the SUT kept busy.DiskBytes the SUT read from and wrote to its block devices.
readwritten
Phase
Starts
Seconds
Operations
ops/s
CPU cores
Peak GiB
Page cache at end GiB
P0 environment check
0 s
0.8
0
0.0
0.00
0.00
0.00
P1 schema definition
1 s
0.7
0
0.0
0.00
0.00
0.00
P2 import
1 s
8.3
0
0.0
0.08
0.03
0.02
P3 cold single-client run
10 s
17.2
135
7.8
0.00
0.01
0.00
P4 warm-up
27 s
11.5
8,720
758.2
0.81
0.06
0.04
P5 warm single-client run
38 s
2.6
280
108.1
0.59
0.05
0.04
P6 read, write and modify ladders
41 s
116.2
86,712
746.1
0.82
0.22
0.12
P7 mixed ladder
157 s
289.0
88,901
307.6
0.47
0.14
0.00
P9 memory floor search
446 s
22.7
8,386
369.7
0.79
0.08
0.06
P11 final footprint and surface audit
469 s
1.4
0
0.0
0.10
0.07
0.06
P12 teardown
471 s
1.5
0
0.0
0.00
0.07
0.06
12. From the cold restart
P3 restarts the SUT with an empty page cache and runs every class from one client. P4 warms the system up under load, and P5 runs every class again. These charts start at the restart, so they show how long the system takes to reach its warm speed and what it holds while it gets there.
Every single-client operationEach dot is one operation, on a log scale. P3 runs every class in turn on an empty page cache; P5 runs them again after the warm-up.
P3 coldP5 warm
ThroughputValid operations per second, ramps and warm-ups included.Latencyp99 of every operation in each second, all classes together, on a log scale.MemoryThe SUT's cgroup memory, stacked: what the processes hold, shared memory such as a buffer pool, and the page cache the kernel can reclaim. The envelope allows 8 GiB.
process (anon)shared memorypage cachekernel
CPUCores the SUT kept busy.DiskBytes the SUT read from and wrote to its block devices.
readwritten
Cold and warm mediansEach class's p50 on an empty page cache (P3) and after the warm-up (P5), log scale. The longer the line, the more the class depends on a warm cache.