Re: [dpdk-dev] [PATCH 0/6] vhost/virtio performance loopback utility

[Thomas Monjalon <thomas.monjalon@6wind.com>]

CC Pablo, testpmd maintainer

Pablo,
This patchset looks really valuable to improve performance debugging.
Would you have time to dig into a review please?


2016-05-05 18:46, Zhihong Wang:
> This patch enables vhost/virtio pmd performance loopback test in testpmd.
> All the features are for general usage.
> 
> The loopback test focuses on the maximum full-path packet forwarding
> performance between host and guest, it runs vhost/virtio pmd only without
> introducing extra overhead.
> 
> Therefore, the main requirement is traffic generation, since there's no
> other packet generators like IXIA to help.
> 
> In current testpmd, io-fwd is the ideal candidate to perform this loopback
> test because it's the fastest possible forwarding engine: Start testpmd
> io-fwd in host with 1 vhost pmd port, and start testpmd io-fwd in the
> connected guest with 1 corresponding virtio pmd port, and these 2 ports
> form a forwarding loop, packets received by the host vhost pmd port are
> forwarded to the guest virtio pmd port, and packets received by the guest
> virtio pmd port are sent to the host vhost pmd port.
> 
> As to traffic generation, "start tx_first" injects a burst of packets into
> the loop, which is the ideal way to do that.
> 
> However 2 issues remain:
> 
>    1. If only 1 burst of packets are injected in the loop, there will
>       almost definitely be empty rx operations, e.g. When guest virtio pmd
>       port send burst to the host, then it starts the rx immediately, it's
>       likely the packets are still being forwarded by host vhost pmd port
>       and haven't reached the guest yet.
> 
>       We need to fill up the ring to keep all pmds busy.
> 
>    2. io-fwd doesn't provide retry mechanism, so if packet loss occurs,
>       there won't be a full burst in the loop.
> 
> To address these issues, this patch:
> 
>    1. Add an io_retry-fwd in testpmd to prevent most packet losses.
> 
>    2. Add parameter to enable configurable tx_first burst number.
> 
> Other related improvements include:
> 
>    1. Handle all rxqs when multiqueue is enabled: Current testpmd forces a
>       single core for each rxq which causes inconvenience and confusion.
> 
>    2. Show topology at forwarding start: "show config fwd" also does this,
>       but show it directly can reduce the possibility of mis-configuration.
> 
>    3. Add throughput information in port statistics display for "show port
>       stats (port_id|all)".
> 
> Finally there's documentation update.