Sequence 09 - 名词解析和缩写速查
Top
首页从 00 开始总纲

Sequence 09 - 名词解析和缩写速查

这一页是你阅读其他文档时的索引页。目标不是背百科，而是让你在面试时能做到：

看到缩写 -> 知道全称 -> 知道它在系统里的位置 -> 知道它和 JD/CV 的关系 -> 能讲一句英文解释

1. 总图：这些词分成哪几类

flowchart TB
    Terms[Interview Terms] --> AI[AI Inference]
    Terms --> GPU[GPU / CUDA]
    Terms --> Net[Networking / Data Movement]
    Terms --> Dist[Distributed Training / Serving]
    Terms --> Perf[Performance / Observability]
    Terms --> Quant[Quant / Trading Systems]
    Terms --> Infra[Infra / Reliability]

    AI --> KV[KV Cache / Prefill / Decode]
    GPU --> CUDA[CUDA / Warp / Stream / Nsight]
    Net --> UCX[UCX / RDMA / GPUDirect / NIXL / GPUNetIO]
    Dist --> NCCL[NCCL / DP / TP / PP / FSDP]
    Perf --> Metrics[TTFT / TPOT / P99 / Throughput]
    Quant --> OMS[OMS / Order Book / Slippage / MTM / PnL]
    Infra --> SLO[SLO / Circuit Breaker / Backpressure]

2. AI inference / LLM serving

缩写/名词 全称 是什么 面试一句话
LLM Large Language Model 大语言模型。面试里不要只说模型，要说 request path、prefill、decode、KV、batching。 An LLM serving system is a runtime and scheduling problem, not just a model call.
Inference 推理 模型在线生成输出的过程。系统重点是 latency、throughput、tail latency、cost。 Inference performance depends on scheduling, memory, communication, and batching.
Serving 模型服务化 把模型变成可在线调用的服务，包括路由、调度、batching、streaming、监控。 Serving turns model execution into a production system.
Prefill 输入上下文计算阶段 对 prompt 做前向计算，建立 KV cache。长 prompt 会拉高 TTFT。 Prefill is compute-heavy and strongly affects time to first token.
Decode 逐 token 生成阶段 每一步生成一个 token，反复读取 KV cache。影响 TPOT。 Decode is latency-sensitive and repeatedly reads KV cache.
KV Cache Key/Value cache Attention 层保存历史 token 的 K/V，减少重复计算，但占显存。 KV cache is inference state and becomes a memory/data-movement problem at scale.
TTFT Time To First Token 从请求进入到第一个 token 输出的时间。 TTFT reflects queueing, scheduling, prefill, and KV allocation.
TPOT Time Per Output Token 输出阶段每个 token 的平均耗时。 TPOT reflects decode loop efficiency.
P99 99th percentile latency 99% 请求比这个值快，1% 更慢；比平均值更能说明线上体验。 P99 shows tail behavior under saturation and interference.
Dynamic batching 动态批处理 把不同请求动态合批，提高 GPU 利用率，但可能影响公平性和尾延迟。 Batching improves throughput but can hurt fairness and P99.
Continuous batching 连续批处理 decode 中动态加入/移除请求，提高在线 serving 利用率。 Continuous batching keeps GPU busy during iterative decoding.
Disaggregated serving 分离式服务 把 prefill 和 decode 放到不同 worker/pool，减少资源干扰。 Disaggregation helps when prefill and decode have different resource profiles.
Speculative decoding 推测解码 用小模型或 draft 模型先生成，再由大模型验证，加速 decode。 Speculative decoding trades extra compute for lower generation latency.

3. GPU / CUDA / profiling

缩写/名词 全称 是什么 面试一句话
CUDA Compute Unified Device Architecture NVIDIA GPU 编程平台。这个 JD 不是纯 CUDA kernel 岗，但要求你懂 GPU 编程模型。 CUDA is the main programming model for NVIDIA GPU compute.
Kernel CUDA kernel 在 GPU 上并行执行的函数。 A kernel is launched by CPU and executed by many GPU threads.
Grid Grid 一次 kernel launch 的所有 block。 A grid is the full parallel work launched for a kernel.
Block Thread block 一组线程，可共享 shared memory。 A block is the unit of cooperative threads on an SM.
Thread GPU thread GPU 上的执行单元。 Threads execute the same kernel over different data elements.
Warp Warp NVIDIA GPU 上通常 32 个线程组成的执行组。 Warp-level behavior matters for divergence and memory access.
SM Streaming Multiprocessor GPU 上执行 thread blocks 的硬件单元。 SMs are where CUDA blocks are scheduled and executed.
Occupancy 占用率 活跃 warp 数相对硬件可支持上限的比例，不是越高越好。 Occupancy helps hide latency but does not guarantee performance.
Coalescing 合并访存 相邻线程访问连续内存，使 memory transaction 更高效。 Coalesced access improves effective memory bandwidth.
Shared memory 共享内存 block 内 on-chip memory，低延迟，常用于 tiling/reuse。 Shared memory reduces repeated global memory traffic.
Pinned memory Page-locked host memory 不可分页 host memory，更适合 DMA 和 async copy。 Pinned memory enables faster and more predictable host-device transfers.
Stream CUDA stream GPU 异步操作队列。 Streams enable overlap between compute and copies.
Event CUDA event GPU 侧计时和同步工具。 Events are used for GPU timing and dependency control.
Nsight Systems NVIDIA timeline profiler 看端到端 timeline：CPU、GPU、copy、sync、idle。 Nsight Systems finds the end-to-end critical path.
Nsight Compute NVIDIA kernel profiler 看单个 kernel 内部指标：memory、warp stalls、occupancy。 Nsight Compute explains why a kernel is slow.
RenderDoc Graphics debugger/profiler Vulkan/WebGPU/graphics workload 分析工具。 RenderDoc experience is GPU-adjacent but not a CUDA substitute.

4. Networking / data movement

缩写/名词 全称 是什么 面试一句话
NIC Network Interface Card 网卡。AI data center 里常关心 NIC/GPU/CPU 拓扑。 NIC placement and topology affect GPU communication performance.
RNIC RDMA-capable NIC 支持 RDMA 的网卡。 RNIC can access registered memory without normal kernel TCP path.
RDMA Remote Direct Memory Access 远端直接内存访问，减少 CPU/kernel involvement。 RDMA reduces CPU overhead and can provide low-latency data movement.
RoCE RDMA over Converged Ethernet 在 Ethernet 上跑 RDMA。需要网络配置支持低丢包/拥塞控制。 RoCE brings RDMA semantics onto Ethernet fabrics.
InfiniBand 高性能互联网络 HPC/AI 集群常用低延迟高带宽网络。NVIDIA/Mellanox 强项。 InfiniBand is a high-performance fabric for HPC and AI clusters.
Ethernet 以太网 通用网络，AI 集群中也可通过 Spectrum-X/RoCE 等增强。 Ethernet is operationally common but needs AI-specific optimization at scale.
UCX Unified Communication X 通信传输抽象，支持 TCP、RDMA、shared memory、CUDA-aware path。 UCX is a transport abstraction layer for high-performance communication.
UCP UCX Protocol layer UCX 面向应用的高层通信 API。 UCP exposes endpoint and request abstractions.
UCT UCX Transport layer UCX 底层 transport API。 UCT maps to low-level transport capabilities.
GPUDirect RDMA GPU Direct RDMA RNIC 直接访问 GPU memory 的 data path。 GPUDirect RDMA avoids staging through CPU memory when supported.
GDR GPUDirect RDMA 同上，常见缩写。 GDR is about GPU memory being reachable by the NIC.
GDS GPUDirect Storage GPU 与存储之间减少 CPU staging 的路径。 GDS targets storage-to-GPU data movement.
NIXL NVIDIA Inference Transfer Library 面向 AI inference 的点对点数据/状态传输库，典型是 KV/state movement。 NIXL is for inference data movement, not a collective library.
GPUNetIO GPU Network IO DOCA 组件，让 GPU 更直接参与 packet/data processing。 GPUNetIO is useful when GPU should consume or process network data directly.
DOCA Data Center-on-a-Chip Architecture NVIDIA DPU/BlueField 软件框架。 DOCA provides programmable infrastructure services around networking and DPUs.
DPU Data Processing Unit 数据中心基础设施处理器，如 BlueField。 DPU offloads networking, storage, and security infrastructure tasks.

5. Distributed AI / communication libraries

缩写/名词 全称 是什么 面试一句话
NCCL NVIDIA Collective Communications Library GPU collective 通信库。 NCCL accelerates collectives such as all-reduce across GPUs.
MPI Message Passing Interface HPC 常用消息传递接口。 MPI is a general distributed communication standard.
Rank 分布式进程编号 每个参与通信的 worker/process/GPU rank。 Rank identifies a participant in distributed computation.
World size 总 rank 数 分布式任务总参与者数量。 World size determines the communication group size.
All-reduce 集合通信 每个 rank 输入，聚合后每个 rank 得到结果。 All-reduce is central to data-parallel gradient synchronization.
All-gather 集合通信 收集所有 rank 的分片到每个 rank。 All-gather is common in tensor parallel and sharded execution.
Reduce-scatter 集合通信 reduce 后把结果分片给不同 rank。 Reduce-scatter reduces and shards the result.
DP Data Parallelism 数据并行，多卡复制模型、分数据。 DP scales batch processing but needs gradient communication.
TP Tensor Parallelism 张量并行，把单层矩阵/attention 切到多 GPU。 TP reduces per-GPU compute/memory but adds collective communication.
PP Pipeline Parallelism 流水线并行，把模型层切到不同 GPU。 PP trades memory capacity for pipeline scheduling complexity.
FSDP Fully Sharded Data Parallel 参数/梯度/优化器状态分片的数据并行。 FSDP reduces memory by sharding model states.
ZeRO Zero Redundancy Optimizer DeepSpeed 的状态分片优化方法。 ZeRO reduces redundant optimizer/model state memory.
algbw Algorithm bandwidth nccl-tests 的算法视角带宽。 algbw reflects useful payload throughput.
busbw Bus bandwidth nccl-tests 对总线流量估算的带宽。 busbw estimates fabric traffic impact.

6. Performance / reliability

缩写/名词 全称 是什么 面试一句话
Throughput 吞吐 单位时间处理量，如 req/s、tokens/s、orders/s。 Throughput must be considered together with latency and saturation.
Latency 延迟 单个请求/操作耗时。 Latency should be broken down by critical path.
Tail latency 尾延迟 P95/P99/P999 等高分位延迟。 Tail latency often exposes queueing and interference.
SLO Service Level Objective 服务目标，如 P99 < 200ms。 SLO defines what performance is acceptable.
SLA Service Level Agreement 对外承诺的服务协议。 SLA is usually stricter because it has business consequences.
Backpressure 反压 下游处理不过来时限制上游输入。 Backpressure prevents overload from becoming cascading failure.
Circuit breaker 熔断器 错误/异常超过阈值时临时阻断调用或交易。 Circuit breakers protect the system under abnormal conditions.
Rate limiter 限流器 限制请求/订单速率。 Rate limiting protects APIs and downstream services.
Retry storm 重试风暴 大量失败重试放大系统压力。 Retries need backoff, jitter, and budget.
Benchmark 基准测试 控制变量测性能。 Benchmarking turns performance claims into evidence.
Profiling 性能剖析 找耗时和资源瓶颈。 Profiling should precede optimization.
Observability 可观测性 metrics/logs/traces/alerts。 Observability makes failures diagnosable.

7. Quant / trading system

缩写/名词 全称 是什么 面试一句话
Quant Quantitative trading 用数据、模型、规则和系统执行交易。 Quant systems combine strategy, data, execution, risk, and accounting.
CeFi Centralized Finance 中心化交易所，如 Binance/OKX/Bitget。 CeFi trading emphasizes exchange connectivity and low-latency execution.
DeFi Decentralized Finance 链上金融协议。 DeFi trading adds wallet, signing, mempool, and on-chain settlement risks.
CEX Centralized Exchange 中心化交易所。 CEX systems require robust market data, order routing, and risk controls.
DEX Decentralized Exchange 去中心化交易所。 DEX execution depends on routing, slippage, gas, and chain confirmation.
OMS Order Management System 订单管理系统，管理订单生命周期和状态机。 OMS is the control plane for orders.
EMS Execution Management System 执行管理系统，负责执行策略、路由、拆单。 EMS turns intent into executable orders.
Order book 订单簿 买卖挂单队列。 Order book depth determines slippage and executable size.
BBO Best Bid and Offer 最优买价和最优卖价。 BBO is the top of book.
Mid price 中间价 (best bid + best ask)/2。 Mid price is a common mark/fair price proxy.
Spread 买卖价差 ask - bid。 Spread is both cost and opportunity.
Slippage 滑点 预期价格和实际成交价格偏差。 Slippage must be modeled before execution.
TWAP Time Weighted Average Price 按时间切片执行/计算的均价。 TWAP reduces timing impact but may miss fast opportunities.
VWAP Volume Weighted Average Price 按成交量加权均价。 VWAP reflects where volume actually traded.
MTM Mark-to-Market 按当前市场价格重估持仓。 MTM prevents stale PnL and hidden risk.
PnL Profit and Loss 盈亏。 PnL must distinguish realized and unrealized components.
FIFO First In First Out 先进先出成本匹配。 FIFO is used to calculate realized PnL.
WAVG Weighted Average Cost 加权平均成本。 WAVG is useful for current position cost and unrealized PnL.
Drawdown / DD 回撤 从权益高点跌下来的幅度。 Drawdown is a key risk constraint.
Pre-trade risk 交易前风控 下单前检查余额、仓位、滑点、限额。 Pre-trade risk must be on the hot path.
Kill switch 紧急停机开关 异常时停止交易/撤单/禁止下单。 A kill switch is mandatory for live trading.
Position 仓位 当前持有资产/合约数量。 Position drives exposure and risk.
Exposure 敞口 对价格波动暴露的风险。 Exposure must be bounded and monitored.
Hedge 对冲 用反向或相关资产降低风险。 Hedging turns directional risk into execution and basis risk.
Market maker / MM 做市商 双边报价提供流动性。 Market making is about quoting, inventory, hedging, and risk.
LP Liquidity Provider 流动性提供者。 LP obligations impose uptime, depth, and spread constraints.
MEV Maximal Extractable Value 交易排序/打包带来的可提取价值。 MEV requires careful simulation and execution protection.
Mempool Memory pool 待打包交易池。 Mempool monitoring can provide early signals but adds noise and risk.
HFT High-Frequency Trading 高频交易。 HFT is mostly about latency, determinism, and risk controls.

8. 面试里怎么用这页

不要把这页背成词典。正确用法：

如果被问 “NIXL 和 NCCL 区别？”
  先定位：NCCL 是 collectives，NIXL 是 inference state/data transfer。
  再讲 data path：NIXL 典型处理 KV/state movement。
  最后讲验证：看 TTFT/P99/KV transfer bandwidth/CPU overhead。

如果被问 “quant 项目和 NVIDIA JD 有什么关系？”
  先定位：它不是 GPU networking 项目。
  再讲迁移能力：hot path、latency、event-driven architecture、risk gate、benchmark、observability。
  最后讲边界：trading execution 的低延迟方法论可迁移，但不是 RDMA/UCX production ownership。
Top