RandomX sustained H/s performance authority

[Deadbytes101]

RIGOS performance phase for sustained RandomX throughput without sacrificing appliance stability.

Physical baseline

• source: 223794509e82d5dedda79da87a11ed2941c16c34
• image SHA-256: 736a265eec2d27c1dc746fb9d5847e76be2e954f0d7d9a041bab1c7bf30963d8
• CPU: Intel Core i3-2310M, 2 cores / 4 threads, 3 MiB L3
• memory: 8 GiB dual-channel DDR3-1333
• miner profile: exact 2 threads
• regular huge pages: 100%, 1168/1168 dataset and 2/2 mining threads
• sustained observed rate: approximately 340.7–341.2 H/s at 60 seconds, max 341.7 H/s
• accepted/rejected observed: 26/0
• package temperature observed: 81 C
• CPU clock under load: approximately 2.093 GHz on all logical CPUs
• governor: schedutil, intel_pstate passive
• MSR module/device: absent
• XMRig assembly report: auto:none
• systemd miner restarts: 0

Performance blockers and open questions

1. RandomX MSR prefetcher optimization is not active.
2. Exact 1/2/3/4-thread sustained benchmark matrix has not been measured.
3. Thermal trip points and throttle behavior need a hard evidence gate before any aggressive policy.
4. Miner health observer reported no_recent_speed_evidence despite recent XMRig speed lines; observability must be corrected before soak scoring.
5. Governor changes must be benchmark-driven; current mining load already holds maximum non-turbo clock.
6. 1 GiB huge pages must not be enabled unless the CPU/kernel explicitly support them.

Required engineering doctrine

• maximize sustained H/s, not startup spikes
• preserve zero rejected shares and bounded restart behavior
• no unsupported MSR writes
• no permanently root-running miner
• snapshot and restore every modified MSR value
• rollback all touched CPUs on partial failure
• temperature and throttle evidence are mandatory
• benchmark changes one variable at a time
• publish exact JSON evidence for every result

Acceptance gates

• controlled MSR authority with CPU allowlist, readback, rollback, and shutdown restore
• 1/2/3/4-thread benchmark matrix with at least 15 minutes per candidate
• sustained 15-minute H/s, temperature, clock, rejected shares, and restart count recorded
• winner selected by sustained H/s subject to thermal and stability gates
• 24-hour single-node performance soak
• no regression in USB/state/internal-disk safety

This issue does not claim HiveOS performance parity or superiority until physical benchmark and soak evidence pass.

[Deadbytes101]

Physical performance truth snapshot

Exact baseline:

• source 223794509e82d5dedda79da87a11ed2941c16c34
• image SHA-256 736a265eec2d27c1dc746fb9d5847e76be2e954f0d7d9a041bab1c7bf30963d8
• Intel Core i3-2310M, 2 cores / 4 SMT threads, 3 MiB L3
• all logical CPUs online; SMT active
• sustained clock approximately 2.093 GHz under mining load
• intel_pstate passive, governor schedutil, turbo disabled
• package temperature 81 C; ACPI zones 80 C
• 8 GiB dual-channel DDR3-1333
• exact 2-thread mining profile
• huge pages 100%; authority allocated 1280 pages
• observed 60-second rate approximately 340.7–341.2 H/s, max 341.7 H/s
• accepted/rejected observed 26/0
• zero systemd miner restarts
• MSR module not loaded; /dev/cpu/0/msr absent
• XMRig reports ASSEMBLY auto:none
• CPU and memory PSI showed no active pressure

Observed correctness defect: miner-health-status.json reported degraded/no_recent_speed_evidence while speed and accepted-share lines were visibly present in the same recent journal window. This must be fixed before performance soak scoring.

Interpretation:

• governor is not the first bottleneck because sustained frequency is already at the hardware ceiling
• regular huge pages are already fully effective
• 1 GiB pages are not authorized without an explicit CPU/kernel support gate
• the first controlled performance candidate is reversible Intel RandomX MSR prefetcher optimization
• thread count still requires a controlled 1/2/3/4 matrix; 2 threads are current baseline, not an assumed optimum

[Deadbytes101]

Implementation branch opened

Draft PR #15: performance/randomx-sustained-hash

Current first batch adds a reversible per-CPU MSR authority with:

• exact CPU allowlist: GenuineIntel family 6 model 42 only
• direct MSR 0x1a4 access with target 0xf
• original-value snapshot before mutation
• readback verification
• idempotent re-entry
• rollback after partial apply failure
• state retention when rollback is incomplete
• restore on service stop
• stale-state discard after reboot
• optional miner dependency through Wants= rather than Requires=
• unprivileged miner preserved
• behavioral source tests and built-image verification

No governor, thread-count, overclock, or 1 GiB-page mutation is included in this batch. The PR remains draft until source verification, exact image verification, and physical before/after evidence pass.

[Deadbytes101]

RandomX MSR authority source verification: PASS

Exact performance head 528c4e4a162089fec8f2568337ec1dda45085b65 passed the full authoritative source gate.

Confirmed:

• reversible MSR authority behavior tests 4/4
• unsupported CPU remains non-mutating
• partial write failure rolls back recoverable CPUs and retains state when incomplete
• supported apply is idempotent and restore returns original values
• Alpha8 builder non-login shell contract remains intact
• existing Hive-exit stability contracts remain green
• full workspace, schema, release, systemd, and RIGOS verification passed

Final marker: RANDOMX_PERFORMANCE_AUTHORITY_VERIFICATION=PASS.

No physical MSR or H/s claim is made yet. Next gate is a fresh exact image build from this head, followed by image-level verification and controlled before/after physical benchmarking.

[Deadbytes101]

MSR miner-start safety gate added

Performance head advanced to c84dee38784becd7cb18fa26b8bc31b5cd655736.

Before building the performance image, the implementation was hardened so an incomplete MSR rollback cannot silently fall through to XMRig startup. The existing unprivileged miner gate now accepts only current, exact, safe authority outcomes and denies stale/malformed/partial mutation states.

The image verifier was also strengthened to prove the built squashfs contains:

• the kernel msr module
• kmod and modprobe
• enabled rigos-randomx-msr.service
• exact authority and miner-gate bytes
• unprivileged miner execution
• production MSR status enforcement markers

The previous source PASS no longer authorizes an image build because the head changed. Required next step: rerun targeted behavior tests and the full authoritative source gate on this exact head.

[Deadbytes101]

MSR miner-start safety authority: source PASS

Exact performance head 3b8f051b521b52e6e606750499701af4c93cde1f passed the complete authoritative source gate.

Confirmed:

• production miner gate enforces current RandomX MSR authority truth
• supported and unsupported safe paths remain eligible to mine
• incomplete rollback is denied before XMRig start
• RandomX MSR behavior suite passed 4/4
• miner stability regression suite passed 2/2
• Alpha8 builder contract, full workspace tests, schema, release build, systemd ordering, and full RIGOS verification passed

Final marker: RANDOMX_MSR_MINER_GATE_AUTHORITY=PASS.

Next authorized gate is a fresh exact image build from this SHA with squashfs-level proof of the kernel MSR module, kmod/modprobe, enabled authority service, exact authority/miner-gate bytes, and unprivileged miner wiring. No physical H/s improvement claim is made yet.

[Deadbytes101]

Performance image build not started: Podman CLI compatibility

Rootful Podman was confirmed, but the host rejected podman build --pull=never because this Podman release expects a boolean --pull value. The command failed with exit 125 before builder image creation.

No source, builder, image, recovery ISO, or manifest evidence changed. The exact source gate remains valid; the exact image gate remains pending.

Retry requirement: use --pull=false or omit the flag, then rerun the complete rootful build and built-image verification from the beginning.

[Deadbytes101]

Performance image build reached post-build verifier; CRLF root cause fixed

The build from 3b8f051b521b52e6e606750499701af4c93cde1f completed through live-build and standard appliance verification. Evidence passed for rootful Podman, privileged mknod, MSR behavior 4/4, release compilation, BIOS/EFI GRUB installation, systemd ordering, USB image verification, and recovery ISO verification.

Only the final RandomX-specific verifier path was malformed. Root cause: version.env was sourced from the Windows working tree and carried CR bytes because it lacked an explicit LF attribute.

New exact head: 96ef72a2483293cc24d734adf027f5223a413830

The entrypoint now reads version authority from the exact Git blob, rejects CR bytes, and .gitattributes pins build/usb/version.env to LF. A regression test locks this behavior.

The generated image from the superseded head is not authorized for physical testing. Next gate is fresh source verification and exact rebuild from 96ef72a....

[Deadbytes101]

Program expansion

The broader goal is now tracked in #16: RIGOS advanced mining architecture program.

Issue #14 remains the physical RandomX sustained-H/s authority for the current i3-2310M target. It should produce the first trustworthy tuning winner and feed evidence into the larger architecture program.

Immediate order remains:

1. exact image from the final PR RandomX sustained-hash performance authority #15 head
2. built-image verification
3. physical MSR apply/readback/restore
4. repair recent-speed observability
5. controlled 1/2/3/4-thread and affinity matrix
6. thermal-gated winner selection
7. 24-hour soak

No generalized competitor superiority claim is authorized before the same-hardware bake-off defined in #16.

[Deadbytes101]

Exact source head authorized for rebuild

Head 96ef72a2483293cc24d734adf027f5223a413830 passed the complete version-authority and RandomX source gate.

Confirmed:

• exact Git-blob version authority
• no CR contamination
• CRLF regression test 1/1
• RandomX MSR behavior 4/4
• full workspace, schema, release, systemd, and RIGOS verification

Final marker: RANDOMX_VERSION_AUTHORITY_FIX=PASS.

Next gate is a fresh exact image rebuild from this SHA. The earlier image produced from 3b8f051... remains superseded and is not authorized for physical testing.

[Deadbytes101]

Kernel MSR image evidence gate refined

The exact image build from 96ef72a... completed through standard appliance verification but stopped at the RandomX kernel support proof. The old verifier only recognized a loadable msr.ko* file and did not recognize kernels that record the MSR driver as built-in through modules.builtin.

New head: bb799c9cb19b2018d8b83ba0789e4e84349774fc.

The verifier now accepts exactly two safe representations:

1. compressed or uncompressed kernel/arch/x86/kernel/msr.ko*
2. an exact kernel/arch/x86/kernel/msr.ko entry in a kernel modules.builtin file

Missing support in both places still fails closed. The verifier emits module or builtin in successful evidence.

Fresh source verification and rebuild are required before physical testing.

[Deadbytes101]

Kernel MSR package authority: PASS

Direct squashfs inspection proved the performance image contains working kernel MSR packaging:

• CONFIG_X86_MSR=m
• exact usr/lib/modules/6.1.0-50-amd64/kernel/arch/x86/kernel/msr.ko
• installed linux-image-6.1.0-50-amd64 and linux-image-amd64
• matching modules.order and DPKG package file-list evidence

The previous built-image rejection was a verifier false-negative caused by grep -q terminating early inside a pipeline while pipefail was enabled.

New head 5281b25931f8636059208fb9ab30f64084346d99 removes quiet grep from those pipelines, adds a regression lock, and executes that regression before the costly USB build.

Next gate: source verification on the new head, then one fresh exact image rebuild and built-image verification. No physical flash is authorized yet.

[Deadbytes101]

RandomX MSR service activation defect fixed

The corrected built-image verifier reached the service-enablement gate and proved the existing diagnostic image does not enable rigos-randomx-msr.service at boot.

New head de6563053b31fa79c77e268623d08b9ff923f62b updates the live-build chroot hook to:

• mark the RandomX MSR authority executable
• enable rigos-randomx-msr.service before rigos-miner.service
• lock both requirements in the build-entrypoint regression test

The older image cannot be promoted. Fresh source verification and an exact rebuild are required before any physical MSR or throughput test.

[Deadbytes101]

Exact source authorized for final rebuild

Head de6563053b31fa79c77e268623d08b9ff923f62b passed the RandomX MSR enablement source gate.

Evidence includes exact hook enablement, regression 1/1, MSR authority 4/4, full workspace verification, release build, systemd ordering, and final marker RANDOMX_MSR_ENABLEMENT_SOURCE_GATE=PASS.

Next gate is one fresh exact image rebuild from this SHA, followed by built-image verification and artifact identity capture. The older image remains diagnostic-only and is not authorized for flashing.

[Deadbytes101]

Exact image authority established

Head de6563053b31fa79c77e268623d08b9ff923f62b produced a fully verified performance image.

• image SHA-256: 9a2100dbb359b70d142e99ea31cd9f0d9b0ed271cf988848ba5c9e8e5b5b72ed
• image size: 2685403136 bytes
• recovery SHA-256: 53d02ffdf85cd2c4ccd2a92ad2884b084f9c3d59feca907ce4b518ec5cc98dce
• recovery size: 772800512 bytes
• kernel MSR support: loadable module
• built-image verifier: PASS
• checksum sidecars: PASS
• manifest identity: PASS
• final marker: RANDOMX_EXACT_IMAGE_BUILD=PASS

Next gate is host-side preflash identity verification, followed by controlled physical MSR/readback and sustained-hash testing. No throughput gain is claimed yet.