如何调整采集间隔到1s？

[qq732552048]

看exporter的间隔写的是60台交换机、900台服务器需要3s的时间，如何优化到1s内完成呢？

[gabrieleiannetti]

在标准情况下，输出器在一台计算机上运行，并管理整个网络。因此，所需的时间较长。然而，可以为每个 "通道适配器（ca）"运行导出器，这样可以大大减少时间。

Question here is, how can the runtime of the exporter be improved...

I would suggest to run the exporter for each CA separately on each host instead of one host for the complete fabric.

@guilbaults Can you please provide an example how to run the exporter for one specific CA with port? I think I am missing something here... I do not get it running with ibquerryerrors for providing an example...

[guilbaults]

Within ibquerryerrors, I think --Ca and --Port is only intended to select the local Infiniband port to use (for setup where there is 2 cards or ports on different networks)

  --Ca, -C <ca>           Ca name to use
  --Port, -P <port>       Ca port number to use

Since ibquerryerrors scan and discover the entire fabric at every execution, I don't think it's possible to make it faster with ibquerryerrors

However, using the C API directly and running those commands in parallel to send the MAD packets to all the switches/ports should be fast enough to get the values in 1 second. With that API, it should also be possible to scan the entire fabric once in a while to discover new switches and nodes (and not at every cycle).

https://github.com/linux-rdma/infiniband-diags/blob/master/src/ibqueryerrors.c#L193
https://github.com/linux-rdma/infiniband-diags/blob/master/src/ibqueryerrors.c#L397

Redoing the exporter in Go and gathering the counters directly without using ibquerryerrors is something I had in mind, this will also fix #48 because it will also report nodes without errors. I currently do not really have time to redo this exporter.

[gabrieleiannetti]

I thought to execute the exporter locally on each host to query its channel adapter.
Then the runtime should be pretty low instead of querying the whole fabric.

I think we can close the issue?

[qq732552048]

谢谢您解答了我的疑惑，我明白怎么做了。