GCRI · Foundry
算力与边缘
Risk analysis needs compute in very different places: a large cloud allocation for a national ensemble, a ministry’s own servers for records that may not leave, a field laptop that keeps a model running when the network fails. This page sets out how compute risk develops, from capacity shortfall and concentration in a few providers to jurisdiction, energy and degraded connectivity, and how cloud, local nodes and edge devices are designed to be scheduled as one capacity.
算力与边缘:风险如何形成。
云端、本地节点和边缘的技术能力。下方档案说明风险如何发展、衡量什么、数据来自何处、如何建模以及由谁决定。
需要时容量失效
算力需求在事件期间达到峰值,而此时恰是共享容量被争用、网络受损、电力不稳之际。队列变长,集合缩小,预报在决策窗口关闭后才到达。
集中度
Many institutions rely on a few cloud providers, regions and chip suppliers. An outage, sanction, price change or supply shortage reaches all of them at once, which makes compute concentration a systemic dependency in its own right.
司法管辖区与数据所在地
Where data may be processed is set by law and contract. Moving analysis to the data keeps records in place, but constrains which methods can run and can still leave results open to inference about the inputs.
断开连接与对账
Edge nodes keep working offline and reconcile later. Conflicting edits, clock skew and stale models accumulate during an outage, and reconciliation after a long disconnection may need human judgement rather than an automatic merge.
如何对计算与边缘计算进行建模、测量和测试。
模型、其假设、数据和不确定性,以科学家或工程师会核查的形式呈现。结构遵循已发表的方法。所取数值仅为示意,不是研究发现。
模型、方程与假设
用于算力和边缘的模型结构、其所作假设以及各项假设如何检验。
Ri = f(Hi, Ei, Vi, Ci)危害、风险暴露、脆弱性和应对能力EAL = ∑k pk · Lkover events k with annual probability pkxt+1 = A xt + B ut + εtA carries the couplings, u the shock, ε the error| 假设 | 为何制作 | 如何检验 |
|---|---|---|
| 在一次事件中,危害与暴露相互独立 | 使损失积分保持可计算 | 与基于联合记录拟合的 copula 进行比较 |
| 耦合在工作点附近是线性的 | 让冲击通过 A 传播 | 在线性范围之外施加扰动并比较 |
| 脆弱性曲线可以在不同地点之间迁移使用 | 很少有本地曲线可用 | 在有观测损失的地方据此进行验证 |
- 输入
- 带溯源信息的版本化数据集
- 成果
- 分布,绝不是单一数字
- 代码
- 连同模型卡和测试一起发布
- 复现
- 相同的数据、代码、环境和随机种子
Compute is designed to serve each domain’s heaviest models.
Compute readiness is designed for the Foundry Build and Readiness series. Energy efficiency is reported with ISO/IEC 30134 indicators; no hardware specifications are summed into performance claims.
- 工作负载画像。
- 效率测量。
衡量什么,以及衡量得如何。
每个指标都注明其单位、频率、背后的来源类型以及数值将附带的证据标签。此处不显示数值:数值由计划依据其自身权限提供。
| 指标 | 单位 | 节奏 | 来源类型 | 标签 |
|---|---|---|---|---|
| Queue wait time | minutes, median and tail | 每日 | Scheduler accounting | 观测 |
| 服务可用性 | 占计划时间的比例 | 每月 | 可用性监测 | 观测 |
| 部署在许可司法管辖区之外的工作负载 | 计数 | per placement | Placement policy logs | 观测 |
| Edge reconciliation conflicts | 每次重新同步的计数 | 每个事件 | 同步日志 | 观测 |
| 每项工作负载的能耗 | kilowatt-hours | 每次运行 | 电力遥测数据或供应方估算 | 估算 |
| Carbon intensity of electricity used | grams CO2 equivalent per kilowatt-hour | 每小时 | Grid operator and provider data | 估算 |
| 提供的容量与分配的容量对比 | 核时或 GPU 时 | 每月 | Offer and allocation register | 已声明 |
| Recovery time in failover exercises | 分钟 | 每次演练 | 演练记录 | 观测 |
证据来自何处。
具名公共数据集与机构记录,各自注明许可类别。依协议获得的来源仅按其条款使用,绝不再次发布。
| 来源 | 它提供什么 | 许可类别 |
|---|---|---|
| TOP500 和 Green500 榜单 | 已公布的大型 系统 | Free, terms of the project |
| Ember 电力数据 | 各国发电结构和电力碳强度 | CC BY 4.0 |
| Copernicus Data Space Ecosystem | Earth observation archives with processing next to the data | 免费开放,须署名 |
| NASA Earthdata Cloud | 云对象存储中的地球科学档案,用于区域内分析 | Free and open, NASA data policy |
| 公有云区域和状态记录 | 区域位置、可用性历史和事件报告 | 提供方条款 |
| Kubernetes 发布和弃用记录 | Version support windows and API removals that affect workload portability | CC BY 4.0 |
由谁决定,以及证据服务于什么。
Foundry 负责准备证据。以下决定由所列机构依其自身规则作出,绝不由平台作出。
| 谁来决定 | 问题 | 本页面服务的证据 |
|---|---|---|
| 数据所有者和数据保护机构 | 数据可在何处处理,以及结果是否可以出境 | Placement logs, the jurisdiction of each tier, federated execution records |
| 项目负责人 | 事件期间如何在分配额度内确定工作负载的优先次序 | 队列状态、剩余配额、截止日期 |
| 基础设施贡献方和运营方 | What capacity to offer and on what terms | 供给登记册、利用率、能源和支持负荷 |
| Ministries responsible for digital infrastructure | 国家 风险分析 在多大程度上依赖外国或集中的提供方 | 服务提供方与区域依赖图;退出与故障切换计划 |
该主题产生的产出
- 按数据类别、司法管辖区、紧迫性和成本确定的放置政策
- Workload packages portable across cloud, private and edge tiers
- 带演练记录的故障切换和降级运行手册
- 已提供、已分配和已使用容量的登记册,分别保存
计算工程师和数据所有者提出的问题。
- Where is the data legally and contractually allowed to be processed?
- Which workloads need burst capacity, and which need steady, local availability?
- 在连接性下降或中断时,工作如何继续?
- 如何区分提供的承保能力与实际分配和使用的承保能力?
横跨三个层级的统一控制平面。
NXS-COR (compute, nodes and runtimes) is designed as the compute control plane across cloud, private infrastructure and edge. Workloads are packaged with their dependencies so they can run on any tier that the project’s policy allows, and the scheduler places them according to data location, sensitivity, cost and urgency. Where data cannot move, the analysis is designed to run where the data is held and only permitted results leave.
Edge capacity matters most when it is hardest to provide: during the events risk work is about. Field kits, observatory nodes managed through NXS-OPS (instruments, sensors and calibration) and offline operation of NXS-NST (civilian common operating picture) are designed so that coordination continues through outages, synchronising when connectivity returns. Energy use, cooling and hardware supply are part of the engineering, and the platform’s computing questions connect to the energy and data-centre analysis at Data and Compute.
哪些已经工程化,哪些依然困难。
Compute to data
Federated execution keeps sensitive data in place and moves code and permitted results.
降级运行
边缘节点在本地继续运行,随后再进行同步,冲突会被标记,而不会被悄然覆盖。
Offer is not allocation
Contributed infrastructure is recorded as an offer until it is separately allocated.
局限:联邦分析限制了可运行的方法,边缘设备能力有限,长时间断连后的对账可能需要人工判断。不将各项硬件规格汇总为性能声明。
如何研究,以及方法无法说明的内容。
Where this page’s references lead.
阅读本页时所对照的机构和文件。状态和版本会变化;每个链接都指向发布机构。
问题与解答。
Why move analysis to the data instead of the reverse?
What happens to edge work during an outage?
我的机构可以提供服务器吗?
是否报告能源使用?
Does membership include compute?
是否有开放的算力分配?
从本页贡献证据。
Reports 是联合体的季刊。您对该主题的了解可以成为下一期中的证据、案例或更正,署名归您,并在属于您本人作品时以 CC BY 4.0 发布。
- 电子邮件网关 GW-GCRI从本页发送证据、更正、案例、方法、审查、翻译或概况投稿。它会带上本页标签送达,由负责的编辑接收。
- ZenodoOpen depositDeposit it and submit to the communityPublish your own record on Zenodo with your ORCID iD, then submit it to the Reports community there. Curators review it; the DOI stays yours.
- 报告GCRI查看所有贡献方式贡献类型、各自的评审方式、您获得的署名、许可协议,以及绝不接受的内容。
切勿发送个人数据、敏感基础设施细节或您无权分享的任何内容。贡献在使用前会经过审查;发送贡献不赋予署名权或编辑权,赞助也绝不能换取研究结论。
法律文件。
本网站上的地图、数据和图表仅作为背景和一般信息展示。它们不是建议、预测或预警;信息的局限见免责声明和数据说明。