| title | devops-net-community | |||
|---|---|---|---|---|
| tags |
|
|||
| created | 2024-11-16 10:58:13 UTC | |||
| modified | 2024-11-16 16:58:32 UTC |
- ip
-
Rx is "receive", i.e. stuff coming in and waiting in queue to be processed.
- Tx is "transmit", i.e. stuff waiting in queue to go out.
-
理论上越底层效率越高?
- 主要是内核支不支持
-
pptp ，l2tp 至少是我十年前的玩法了，基本上部署出来直接秒封，流量特征太明显，换 IP 也没用。后面 shadowsockets 也不太行，真正抗的时间久的还是 trojan。
-
经验来看，这些大部分都很容易被封，特别是古老的pptp、l2tp这类。ikev2、vpn这些一旦频率过高也会封。
-
l2tp/ipsec 数据层就是走的 gre 封装
-
https://x.com/hubingkang/status/1900362862767595638
sudo lastb | wc -l
-
紧查了查，我的几台国外服务器，20多万次。国内的反而少一点，不到两万次。但是我内网穿透的家里主机，竟然也有将近2万次，家里的电脑密码很简单，也不知被攻破了没
- 家里内网得注意，一被破解了不是盗取资料就是要勒索了。
-
慌什么，用密钥登录他又无法攻破
- 有些用密钥，有些用的密码
-
用fail2ban拦截
- 还可以自动报 http://blocklist.de，顺便做个 crontab 每天拉那个网站的 ip 黑名单
-
更适合中国宝宝体质的 DNS 测速工具：https://github.com/xxnuo/dns-benchmark/tree/master
-
- 主要解决 AI 辅助网络排障和安全分析时，缺乏底层数据包操作能力的痛点。底层基于真实的 tshark / Wireshark 套件，你可以直接把 .pcap 文件丢给 CodeX 或 Claude，
- 它会自动提取数据帧，基本覆盖 Tshak 全部功能，已经精简过Tool Calling 的上下文，可用于应急响应分析已经 CTF 竞赛
-
Sandbox:
- Coldstart: ~200ms
- Network RTT: ~15ms
- Tool (external): ~20ms
-
LLMs time-to-token:
- Cheap models: ~500ms
- Expensive models: ~1s
- LLM gateways add significant latency
-
Significance of each of these –
- Coldstarts → Impacts your initial load time.
- Network RTT → Minimum latency for any interaction with what's running in the sandbox (e.g. preview deployments, browser use, etc)
- Tool (external) → If running agent outside of the sandbox, this adds to you latency. 15 ms is negligible compared to inference.
- Time-to-token → How long before your agent can start doing something. Larges contributor to latency.
- OpenRouter -> LLM gateways are required if you're passing an LLM token to a sandbox in order to enable per-tenant billing. This latency difference is brutal.
- Be weary of LLM gateways.
-
highly recommend reading the original Latency Numbers Every Programmer Should Know: