菜单
NIC IRQ 1:1 pinning, NVMe queue limits, RPS/XPS, NIC ring buffer, sysctl (UDP/TCP buffers, swappiness, netdev_max_backlog), THP off, and ulimit nofile=1M for Solana / Shredstream nodes. Reboot is required when NVMe queue parameters change.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | slv-irq-tuning |
| description | NIC IRQ 1:1 pinning, NVMe queue limits, RPS/XPS, NIC ring buffer, sysctl (UDP/TCP buffers, swappiness, netdev_max_backlog), THP off, and ulimit nofile=1M for Solana / Shredstream nodes. Reboot is required when NVMe queue parameters change. |
Removes NIC IRQ imbalance and applies Solana-friendly kernel parameters. Runs as part of slv v init / slv r init after the SSH connection check, alongside SMT disable and performance boost.
| # | Item | Effect | Reboot |
|---|---|---|---|
| 1 | NIC IRQ 1:1 pinning | Distributes NET_RX softirq evenly across all CPUs | No |
| 2 | RPS/XPS enable | Software-level packet distribution | No |
| 3 | NIC ring buffer → 8192 (max) | Prevents packet drops during bursts | No |
| 4 | NVMe queues = ncores | Stops queues from being assigned to offline CPUs | Yes |
| 5 | sysctl: UDP/TCP buffers 128MB | Prevents drops on Shredstream / gRPC | No |
| 6 | sysctl: vm.swappiness → 1 | Avoids swapping under memory pressure | No |
| 7 | sysctl: netdev_max_backlog → 50000 | Prevents NIC → kernel queue overflow | No |
| 8 | THP (Transparent Huge Pages) → never | Avoids latency spikes | No |
| 9 | ulimit nofile → 1000000 | Prevents fd exhaustion | No (new sessions only) |
ansible/cmn/irq_tune.yml — applies items 1-9 aboveansible/cmn/files/vs2-irq-tune.sh — re-applies NIC IRQ pin + RPS/XPS + ring + THP at boot (systemd oneshot)ansible/cmn/files/vs2-irq-tune.service — systemd unit for the aboveansible/cmn/files/99-vs2-solana.conf — sysctl parametersansible/cmn/files/99-vs2-solana-limits.conf — /etc/security/limits.d/ansible/cmn/files/vs2-limits.conf — systemd DefaultLimitNOFILE drop-inNormally invoked through cmn/optimize_node.yml together with SMT disable and CPU boost. Standalone:
ansible-playbook -i inventory.yml cmn/irq_tune.yml --limit <host>
Only sets need_reboot=true when NVMe queue parameters in GRUB had to change. The orchestrator reboots the host when needed.
| Check | Expected |
|---|---|
| NIC IRQ queue N → CPU N | 1:1 mapping |
| RPS (bond0/vlan) | ffff etc. |
| NIC ring buffer | RX: 8192, TX: 8192 |
| NVMe IRQ on offline CPU | none |
| net.core.rmem_max / wmem_max | 134217728 |
| net.core.netdev_max_backlog | 50000 |
| vm.swappiness | 1 |
| THP | [never] |
GRUB has nvme.io_queues=N | physical core count |
vs2-irq-tune.service | active (exited) |
| irqbalance | inactive (disabled) |
sudo systemctl disable vs2-irq-tune.service
sudo rm /etc/systemd/system/vs2-irq-tune.service /usr/local/bin/vs2-irq-tune.sh
sudo systemctl enable irqbalance && sudo systemctl start irqbalance
sudo rm /etc/sysctl.d/99-vs2-solana.conf /etc/security/limits.d/99-vs2-solana.conf
sudo rm /etc/systemd/system.conf.d/vs2-limits.conf
sudo systemctl daemon-reexec
sudo sysctl --system
sudo sed -i 's/ nvme.io_queues=[0-9]* nvme.write_queues=[0-9]* nvme.poll_queues=[0-9]*//' /etc/default/grub
sudo update-grub
sudo rm /var/lib/slv/.irq_tuned
# reboot to drop NVMe GRUB params
A small Rust JSON-RPC + WebSocket gateway that fronts yellowstone-faithful (of1) and routes `jet*` analytics methods to ClickHouse (jetstreamer data). Standard Solana RPC methods pass straight through; new analytic methods are answered from precomputed jetstreamer aggregates.
Deploy a self-hosted Pyth Hermes price feed API stack (NATS + Beacon + Hermes) on a VPS. Hermes serves Pyth Network's REST/WS price feeds (`/v2/updates/price/latest`, `/v2/price_feeds`, `/ws`) by listening to Pythnet for accumulator messages and to the Wormhole network (via Beacon, a highly-available spy) for VAAs. The recipe builds all three components from source, drops systemd units, and validates `/v2/price_feeds` returns 200.
Deploy an OSS-clean nftables ruleset to any slv-managed host (Hermes, Pythnet RPC, validator, RPC, …). Inventory-driven — mgmt IPs, public ports, and per-port allow lists all come from `~/.slv/inventory.{network}.{role}.yml`, nothing is hardcoded. Atomic + rollback-safe (failed validation restores the previous config) and persistent across reboot. Used by `slv hermes firewall` and `slv pythnet firewall`; the same `cmn/deploy_nftables.yml` playbook works against any inventory.
Run a self-hosted Pythnet RPC node — the application-specific Solana fork operated by Pyth's data providers. Pythnet RPC is what `slv-hermes` reads from for accumulator messages; running it locally completes a fully self-hosted Pyth price feed stack. Recipe handles disk mount, sysctl tuning, Solana 1.14 fork build (with cstdint workaround for modern GCC), identity gen, and systemd. Snapshot fetch + catch-up typically finishes in 10-20 minutes.
Build, configure, and run anza-xyz/jetstreamer + ClickHouse on a dedicated analytics node. Streams Old Faithful CAR archives at multi-million-TPS into ClickHouse so historical Solana data is queryable via SQL (program_invocations, slot_status, pubkey_mentions). Includes hourly backfill timer and a rolling-window retention option.
Update the kernel (Ubuntu HWE 6.14+), set `amd_pstate=active` (AMD), and apply cpufreq governor=performance, EPP=performance, cpuidle disable, and `scaling_max_freq=cpuinfo_max_freq` so the CPU stays at maximum boost. AMD nodes on kernel < 6.14 require HWE install + reboot.