扫一扫
分享文章到微信
扫一扫
关注官方公众号
至顶头条
最近发现核心交换机有几个下联二层交换机的端口出现了很多丢包,从监控平台上显示如下:
显示为7609的receive discards,登陆上交换机查看端口如下:
从上图中看到有很多数据包overrun(2294627),说明此端口的buffer已经耗尽,其将此buffer里的数据送往PFC的速度慢于此端口的接收数据报的速度,因此这些数据报就会丢弃,如上图Input queue有2289265 drops。overrun的官方解释如下:
Q. What are overruns on a serial interface?
A. Overruns appear in the output of the show interface Serial 0 command when the serial receiver hardware is unable to hand received data to a hardware buffer because the input rate exceeds the receiver's ability to handle the data.
This occurs due to a limitation of the hardware. Overruns occur when the internal First In, First Out (FIFO) buffer of the chip is full, but is still tries to handle incoming traffic. The serial controller chip has limited internal FIFO.
Some chips, for example, have only 256 bytes of buffer space. Data from the network is received into the buffer, whereupon the chip attempts to move the data from the buffer to the router's shared memory for the CPU to process. If the chip is not able to move the data from its internal FIFO buffer into shared memory faster than the rate at which data is received on the interface, then the internal FIFO buffer is full, incoming data is dropped, and the overrun counter is incremented.
------------摘自http://www.cisco.com/en/US/tech/tk801/tk36/technologies_q_and_a_item09186a008014f919.shtml
而同时由上图可以看出此端口没有启用flow-control特性,flow-control的官方解释如下:
因此启用flow-control特性可以避免overrun的现象。但是从根本上来考虑此端口为GE端口速率为1000Mbps,且为full-duplex。而且当时此端口的input速率为22.8Mbps,远没有达到GE的上限。而从硬件架构来看,也远没有达到瓶颈。下为硬件参数显示信息:
查看了一下此板卡WS-X6548-GE-TX的结构图,表示每连续8个GE端口共享1G,而且此8GE端口共享16KB的RX BUFFER和1M的TX/RX BUFFER。所以从结构上来看,应该没有达到性能瓶颈的。
补充:
本周通过与cisco case中心交涉,提交几个show命令显示结果给case中心,然后在case中心的协助下:完成以下2种测试:
1,disable head of line blocking which will utilize the interface buffers instead of the
shared buffers. This will result in only the single over utilized port having drops。查看了一下cisco网站,意思就是说启用head of line blocking使得此8个一组端口的每个端口都启用自己的32kb buffer,而不使用共享的1Mb buffer,这样就能看出到底是哪个端口overrun比较多。详见http://www.cisco.com/en/US/products/hw/switches/ps700/products_tech_note09186a00801751d7.shtml
配置命令如下:
6500(config)#service internal //此命令表示可以启用内部命令,注意内部命令TAB或者?都无效,直接敲入就是了。
6500(config)#interface gigabit 1/1
6500(config-if)#hol-blocking disable
%HOL Blocking is Disabled on: Gi1/1 Gi1/2 Gi1/3 Gi1/4 Gi1/5 Gi1/6 Gi1/7 Gi1/8
这样可以找出来到底是哪一个端口过多overrun,然后将此端口挪到另外一组去。
2,try to config "hold queue 4096 in" to raise the hold queue. please refer to :
http://www.cisco.com/en/US/docs/ios/12_3/interface/command/reference/int_d1g.html#wp1142192
在软件队列处理上通过“hold queue 4096 in”将端口的hold queue调高,看看是否有效果。实际上没有多大效果,依然有很大的丢包。
综合以上测试,case中心给出答复如下:
丢包现象为WS-X6548-GE-TX板卡性能瓶颈所致,即WS-X6548-GE-TX的端口buffer太小,请换成6148A或者6748板卡。下面比较一下此三个板卡的硬件架构:
1),X6548-GE:每8个端口(1-8,9-16,17-24...)共享16KB RX buffer和1MB的RX/TX buffer,并且每8个端口共享1GE带宽,此板卡为fabric-enabled,可以与引擎通过一个8GE的CrossBar或者32GE的系统共享Bus相连;
2),X6148A-GE:每8个端口共享160KB RX buffer,每一个端口独享5.2MB的TX buffer,并且每8个端口共享1GE带宽,但是此板卡为nonfabric-enabled,只能通过32GE的系统共享Bus相连;
3),X6748-GE:每个端口独享166KB的RX buffer和1.164MB的TX buffer,并且此板卡为fabric-enabled,可以与引擎通过每24个端口(1-24,25-48)共享20GE的CrossBar相连或者48GE端口共享32GE的系统共享Bus相连。
本文出自 “Network Manage Tech Experience” 博客
如果您非常迫切的想了解IT领域最新产品与技术信息,那么订阅至顶网技术邮件将是您的最佳途径之一。
现场直击|2021世界人工智能大会
直击5G创新地带,就在2021MWC上海
5G已至 转型当时——服务提供商如何把握转型的绝佳时机
寻找自己的Flag
华为开发者大会2020(Cloud)- 科技行者