[dpdk-dev] DPDK Latency Issue

[dpdk-dev] DPDK Latency Issue

[Jun Han]

Hi all,

I measured a roundtrip latency (using Spirent traffic generator) of sending
64B packets over a 10GbE to DPDK, and DPDK does nothing but simply forward
back to the incoming port (l3fwd without any lookup code, i.e., dstport =
port_id).

However, to my surprise, the average latency was around 150 usec. (The
packet drop rate was only 0.001%, i.e., 283 packets/sec dropped) Another
test I did was to measure the latency due to sending only a single 64B
packet, and the latency I measured is ranging anywhere from 40 usec to 100
usec.

I do not understand why this is so slow. Can someone explain the reasoning
behind this phenomenon?

Thank you so much!

Re: [dpdk-dev] DPDK Latency Issue

[Shaw, Jeffrey B]

Hello,

> I measured a roundtrip latency (using Spirent traffic generator) of sending 64B packets over a 10GbE to DPDK, and DPDK does nothing but simply forward back to the incoming port (l3fwd without any lookup code, i.e., dstport = port_id).
> However, to my surprise, the average latency was around 150 usec. (The packet drop rate was only 0.001%, i.e., 283 packets/sec dropped) Another test I did was to measure the latency due to sending only a single 64B packet, and the latency I measured is ranging anywhere from 40 usec to 100 usec.

40-100usec seems very high.
The l3fwd application does some internal buffering before transmitting the packets.  It buffers either 32 packets, or waits up to 100us (hash-defined as BURST_TX_DRAIN_US), whichever comes first.
Try either removing this timeout, or sending a burst of 32 packets at time.  Or you could try with testpmd, which should have reasonably low latency out of the box.

There is also a section in the Release Notes (8.6 How can I tune my network application to achieve lower latency?) which provides some pointers for getting lower latency if you are willing to give up top-rate throughput.

Thanks,
Jeff

Re: [dpdk-dev] DPDK Latency Issue

[Jun Han]

Thanks a lot Jeff for your detailed explanation. I still have open question
left. I would be grateful if someone would share their insight on it.

I have performed experiments to vary both the MAX_BURST_SIZE (originally
set as 32) and BURST_TX_DRAIN_US (originally set as 100 usec) in l3fwd
main.c.

While I vary the MAX_BURST_SIZE (1, 8, 16, 32, 64, and 128) and fix
BURST_TX_DRAIN_US=100 usec, I see a low average latency when sending a
burst of packets less than or equal to the MAX_BURST_SIZE.
For example, when MAX_BURST_SIZE is 32, if I send a burst of 32 packets or
less, then I get around 10 usec of latency. When it goes over it, it starts
to get higher average latency, which make total sense.

My main question are the following. When I start sending continuous packet
at a rate of 14.88 Mpps for 64B packets, it shows consistently receiving an
average latency of 150 usec, no matter what MAX_BURST_SIZE. My guess is
that the latency should be bounded by BURST_TX_DRAIN_US, which is fixed at
100 usec. Would you share your thought on this issue please?

Thanks,
Jun


On Thu, May 22, 2014 at 7:06 PM, Shaw, Jeffrey B
<jeffrey.b.shaw@intel.com>wrote:

> Hello,
>
> > I measured a roundtrip latency (using Spirent traffic generator) of
> sending 64B packets over a 10GbE to DPDK, and DPDK does nothing but simply
> forward back to the incoming port (l3fwd without any lookup code, i.e.,
> dstport = port_id).
> > However, to my surprise, the average latency was around 150 usec. (The
> packet drop rate was only 0.001%, i.e., 283 packets/sec dropped) Another
> test I did was to measure the latency due to sending only a single 64B
> packet, and the latency I measured is ranging anywhere from 40 usec to 100
> usec.
>
> 40-100usec seems very high.
> The l3fwd application does some internal buffering before transmitting the
> packets.  It buffers either 32 packets, or waits up to 100us (hash-defined
> as BURST_TX_DRAIN_US), whichever comes first.
> Try either removing this timeout, or sending a burst of 32 packets at
> time.  Or you could try with testpmd, which should have reasonably low
> latency out of the box.
>
> There is also a section in the Release Notes (8.6 How can I tune my
> network application to achieve lower latency?) which provides some pointers
> for getting lower latency if you are willing to give up top-rate throughput.
>
> Thanks,
> Jeff
>

Re: [dpdk-dev] DPDK Latency Issue

[Jun Han]

Hi all,

I realized I made a mistake on my previous post. Please note the changes
below.

"While I vary the MAX_BURST_SIZE (1, 8, 16, 32, 64, and 128) and fix
BURST_TX_DRAIN_US=100 usec, I see a low average latency when sending a
burst of packets greater than the MAX_BURST_SIZE.
For example, when MAX_BURST_SIZE is 32, if I send a burst of 32 packets or
larger, then I get around 10 usec of latency. When the burst size is less
than 32, I see higher average latency, which make total sense."


On Mon, May 26, 2014 at 9:39 PM, Jun Han <junhanece@gmail.com> wrote:

> Thanks a lot Jeff for your detailed explanation. I still have open
> question left. I would be grateful if someone would share their insight on
> it.
>
> I have performed experiments to vary both the MAX_BURST_SIZE (originally
> set as 32) and BURST_TX_DRAIN_US (originally set as 100 usec) in l3fwd
> main.c.
>
> While I vary the MAX_BURST_SIZE (1, 8, 16, 32, 64, and 128) and fix
> BURST_TX_DRAIN_US=100 usec, I see a low average latency when sending a
> burst of packets less than or equal to the MAX_BURST_SIZE.
> For example, when MAX_BURST_SIZE is 32, if I send a burst of 32 packets or
> less, then I get around 10 usec of latency. When it goes over it, it starts
> to get higher average latency, which make total sense.
>
> My main question are the following. When I start sending continuous packet
> at a rate of 14.88 Mpps for 64B packets, it shows consistently receiving an
> average latency of 150 usec, no matter what MAX_BURST_SIZE. My guess is
> that the latency should be bounded by BURST_TX_DRAIN_US, which is fixed at
> 100 usec. Would you share your thought on this issue please?
>
> Thanks,
> Jun
>
>
> On Thu, May 22, 2014 at 7:06 PM, Shaw, Jeffrey B <jeffrey.b.shaw@intel.com
> > wrote:
>
>> Hello,
>>
>> > I measured a roundtrip latency (using Spirent traffic generator) of
>> sending 64B packets over a 10GbE to DPDK, and DPDK does nothing but simply
>> forward back to the incoming port (l3fwd without any lookup code, i.e.,
>> dstport = port_id).
>> > However, to my surprise, the average latency was around 150 usec. (The
>> packet drop rate was only 0.001%, i.e., 283 packets/sec dropped) Another
>> test I did was to measure the latency due to sending only a single 64B
>> packet, and the latency I measured is ranging anywhere from 40 usec to 100
>> usec.
>>
>> 40-100usec seems very high.
>> The l3fwd application does some internal buffering before transmitting
>> the packets.  It buffers either 32 packets, or waits up to 100us
>> (hash-defined as BURST_TX_DRAIN_US), whichever comes first.
>> Try either removing this timeout, or sending a burst of 32 packets at
>> time.  Or you could try with testpmd, which should have reasonably low
>> latency out of the box.
>>
>> There is also a section in the Release Notes (8.6 How can I tune my
>> network application to achieve lower latency?) which provides some pointers
>> for getting lower latency if you are willing to give up top-rate throughput.
>>
>> Thanks,
>> Jeff
>>
>
>