Reproducibility appendix

Test environment

The hardware, network, software, and process topology shared by all four provider runs.

Collected after the benchmark suite · no continuous host telemetry was captured

Benchmark host

Hardware and OS facts were collected after all runs. Capacity and topology are stable machine facts; available memory, free disk, transient CPU load, and interface utilization are intentionally not reported as if they were continuous measurements.

CPU	; ; up to 5.485 GHz
CPU cache
ISA highlights
RAM / swap	GiB RAM; GiB swap
Local storage
Data filesystem
OS storage
Physical NIC
Internet channel
Operating system
Host context
Toolchain
Benchmark commit	`fdd0e92b8914f9eda5c0e5c18a0eb7105cbef5e0`

Network distinction. The Intel adapter negotiated a 25 Gbit/s physical link. The operator-stated internet channel available to these remote tests was fixed at 3 Gbit/s. The report does not substitute the NIC link rate for the external bandwidth cap.

Process and cache topology

The benchmark models process and cache isolation on one bare-metal host. It does not reproduce independent-machine CPU, NVMe, NIC, or network-failure domains.

durable dataset                    metadata pressure
+--------------------------+       +--------------------------+
| 16 dataset writers       |       | 1 / 16 / 32 / 64 WALs   |
| 20 GiB RAM / 100 GiB NVMe|       | independent CAS chains   |
+------------+-------------+       +-------------+------------+
             +---------------------------+       |
                                         v       v
                         +-------------------------------+
                         | S3-compatible provider        |
                         +-------------------------------+

payload read sequence

[writer-head: 20 GiB RAM / 100 GiB NVMe] exits
                         |
                         v
[reader-head] [replica 1] [replica 2] run concurrently
 each: 4 GiB RAM / 20 GiB NVMe, 16 read threads

live-follow sequence

16 writer shards -> authoritative WAL CAS -> provider
       local wakeup hints + periodic conditional GET
                         -> 10 logical followers

Workload constants

Profile	Full, schema 4
Payload matrix	4 KiB, 1 MiB, and 16 MiB; 12,044 objects per cohort
Logical dataset	Approximately 200 GiB per provider
Warmup / measurement	15 seconds / 60 seconds per timed size phase
Dataset writers	16
WAL pressure	1, 16, 32, and 64 independent WAL objects
Followers	10 per scenario; 50, 100, 250, 500, and 1,000 ms poll intervals; hint-assisted and poll-only
Correctness	CAS races, durable verification, follower convergence, compaction accounting, and orphan inventory

Privacy and provenance

The public report intentionally omits bucket names, account-bearing endpoints, public and private IP addresses, MAC addresses, and NVMe serial numbers. The old docs/bench-runs/run-1.md informed this page’s presentation only; its earlier host and benchmark facts are not used here.