大网数据广告活动
服务器运维问答

Proving packets per second bottleneck?

· 52CE 我爱测

I have a '100Mb' network connection that's currently consistently transmitting at about 20K packets/sec, irrespective of packet size in the range of 300-600 bytes. This yields an observed bandwidth of 25-98Mb. I'm constantly bein…

问题原文

原帖链接 · user2702772 · CC BY-SA 3.0

I have a '100Mb' network connection that's currently consistently transmitting at about 20K packets/sec, irrespective of packet size in the range of 300-600 bytes. This yields an observed bandwidth of 25-98Mb. I'm constantly being told that because we've not hit the bandwidth limit, we don't have a line problem. I don't agree.

This connection is, on average, running at 60% of maximum the theoretical PPS rate for a 100Mb (copper ethernet) line, once packet size is accounted for. (Although the 100Mb bottle neck is fibre of unknown type, so may have different impact, I don't think that any fibre protocol is better than copper with interpacket gap).

My problem is - without access to the routers or fibre hardware (3rd party provided, can't be helped) how can I prove that we are packet limited? Ideally without causing a massive outage in the process :)

已采纳解答

原帖链接 · Petr Chloupek · CC BY-SA 3.0

Collect the traffic with tcpdump or a similar tool and make a graph of the packet count per time unit. If your assumption is correct, you should see a clear ceiling for the packet count.

You may simulate a counter example by generating many large packets with something like ping -s 1472 -f, it may cause a small outage, so maybe do not do it during the traffic peaks. But 30 seconds may be acceptable for solving a larger problem - you decide.

A switch can be easily the bottleneck as well. Especially cheaper one or a black box router. This was the most common case for a WAN network I was working on. The minimum standard for this kind of traffic was a HP ProCurve line thing. Even an old Cisco was fine as well. But you have to test it.

Also good to mention that among ISPs we in generally used a rule of thumb that 60% utilized line was a fully saturated line. The reason why is that the saturation is basically average over some time of period. But on a shorter time frame you may have overloads by attempting to send just too many packets in the exact same moment which will lead to longer latency. Measure the latency as well. Wireshark is a good tool for a quick analysis like this.

Last but not least, I have not seen any kind of traffic which can fully saturate the line but the ping -s 1472 -f on an otherwise empty line. Once you have multiple connections, you have inefficiencies which lead to lower utilization. Basically, 100Mbit is a theoretical limit under ideal conditions. So the line provider may be right as well and upgrading the line may be the proper solution.

摘自公开归档,保留原文语言,版式经整理,外部媒体与邮箱地址已移除。

资料来源与适用范围

资料来源
Server Fault
适用范围
以原帖系统、软件版本与发布日期为准(2016-05-04)
内容语言
中文;命令、路径和错误标识保留技术原文
本地收录日期
2026-11-15
授权说明
CC BY-SA 3.0

正文与所需资源由本站本地存储。查看来源及更新记录 ↗

请按适用版本和实际日志确认原因;本站没有对每一种系统或设备逐项执行文中的操作。