DataHub 集成 Apache Pulsar：基于 Pulsar Admin API 的 Topic 与 Schema 元数据采集指南
2026/9/20 11:44:20
<op-name>harness run — session<session-id>
【免费下载链接】pypto-gymPyPTO-Gym 是基于 PyPTO 编程框架构建的算子与模型样例仓库项目地址: https://gitcode.com/cann/pypto-gym
Shape:1 orchestrator +<N>subagents · Stage<first>→<last>· wall<wall>· active<active>.
1 · Overall costs
| Metric | Value |
|---|---|
| Wall / active | <wall>run-span ·<active>active (reason<r>· tools<t>· setup<s>) |
| Tokens | <total>= input<in>· cache-rd<crd>· cache-wr<cwr>· output<out> |
| File interactions | <rd>r /<wr>w /<ed>e (+<add>/-<del>ln) over<files>files |
| Tool calls | <n>· top tool<tool>=<share>of tool-exec |
| Generation | <segs>segments ·<blocks>persisted thinking blocks |
| Excluded non-work | <user_wait>user_wait +<idle>idle +<dispatch_wait>dispatch |
2 · Per-stage costs
| Stage | Reason | Tools | Wall | Active | input | cache | output | total | rd/wr/ed |
|---|---|---|---|---|---|---|---|---|---|
<n · name> | <r> | <t> | <wall> | <active> | <in> | <cache> | <out> | <total> | <r/w/e> |
| … |
(按 tools + reason 从重到轻排序各阶段)
3 · Per-agent costs
| Agent / stage | Wall | Active | input | cache | output | total | rd/wr/ed |
|---|---|---|---|---|---|---|---|
<label> | <wall> | <active> | <in> | <cache> | <out> | <total> | <r/w/e> |
| … |
(只列成本最高的少数几个代理——不是全部 N 个)
4 · Reasoning
| Agent / stage | Wall | Active | Reason | segs | output | total |
|---|---|---|---|---|---|---|
<label> | <wall> | <active> | <reason> | <segs> | <out> | <total> |
(推理最重的代理；output token 是推理的产物)
5 · Tool calls
| Agent / stage | Wall | Active | Tool-exec | calls | total | top tools |
|---|---|---|---|---|---|---|
<label> | <wall> | <active> | <toolexec> | <calls> | <total> | <tools> |
运行级：top tool<tool>=<share>of tool-exec.
6 · File interactions
| Agent / stage | Wall | Active | rd | wr | ed (±lines) | files | total |
|---|---|---|---|---|---|---|---|
<label> | <wall> | <active> | <rd> | <wr> | <ed> (+<a>/-<d>) | <files> | <total> |
Most-touched:<file = Nr/Mw/Ke>.
7 · Agent lifecycle
| Agent / stage | Mode | Dispatch | Self-wall | Overhead | Active | total |
|---|---|---|---|---|---|---|
<label> | <async|sync> | <dispatch> | <self-wall> | <overhead|n/a> | <active> | <total> |
Dispatch:<all N async (orchestrator never blocked) | sync overhead Σ <x>>.
Analysis & recommendations
- Bottleneck:
<stage / agent / tool that dominates active time, with the number>. - Tokens:heaviest agent/stage
<= tokens>; cache-rd<crd>is cumulative re-reads — fresh work is input+output<in+out>. - File churn:most-touched
<file = Nr/Mw/Ke>— a file re-read or edited many times signals context churn or a hot spot (e.g. an operator'sMEMORY.mdupdated across every stage). - Wall vs. active:
<wall>run-span but only<active>of real work — the rest was<user_wait + idle>of human/idle time. - Recommendation:
<one or two concrete, contenteditable="false">【免费下载链接】pypto-gymPyPTO-Gym 是基于 PyPTO 编程框架构建的算子与模型样例仓库项目地址: https://gitcode.com/cann/pypto-gym
创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考