[dpdk-users] Significant performance penalty when using flow director in Intel 82599ES

[dpdk-users] Significant performance penalty when using flow director in Intel 82599ES

[Hugo Menna Barreto]

Hi,

I am using an Intel 82599ES NIC with DPDK. Without activating flow director
I am able to process 64B packets at line rate (10Gbps). Unfortunately, as
soon as I activate flow director, the maximum throughput is around 5Gbps --
above this rate the NIC starts dropping packets.

Trying to isolate the problem I inspected the struct obtained using
"rte_eth_stats_get". Interestingly "rte_eth_stats.imissed" is going up
while "rte_eth_stats.rx_nombuf" remains at zero. As far as I know "imissed"
is linked to the MPC register and, therefore, indicates that the NIC buffer
is full.

I can imagine that activating flow director may impact performance -- as it
reduces the effective rx packet buffer size from 512KB to 448KB. However, I
did not expect the impact to be of such proportions. Therefore I am
suspicious that I may be doing something wrong.

My questions are:
When using flow director, is only getting half the throughput the expected
behavior? If it is, is there any direction one should follow in order to --
at least partially -- mitigate this problem? (e.g., would RSC help in this
case?) If it isn't, any ideas so far on what I may be doing wrong?

Thank you for your attention.

Best regards,
Hugo

p.s.

Some relevant fields of "rte_eth_conf":
-------------------------------------
rte_eth_conf {
  [...]
  .rxmode = {
    .mq_mode = ETH_MQ_RX_NONE
    [...]
  }
  .fdir_conf = {
    .mode = RTE_FDIR_MODE_SIGNATURE,
    .pballoc = RTE_FDIR_PBALLOC_64K,
    .status = RTE_FDIR_REPORT_STATUS_ALWAYS,
    [...]
  }
}

Some extra details:
-------------------
* DPDK version: 17.08.1
* NIC: Intel 82599ES (ixgbe)
* CPU: E5-2650 (Sandy Bridge)
* PCIe: Gen2, Speed 5GT/s, Width x8