[dpdk-dev] [PATCH RFC 06/11] mbuf: replace data pointer by an offset

[Venkatesan, Venky]

Olivier, 

The impact isn't going to be felt on the driver quite as much (and can be mitigated) - the driver runs a pretty low IPC (~1.7) compared to some of the more optimized code above it that actually accesses the data. The problem with the dependent compute is like this - in effect you are changing 

struct eth_hdr * eth = (struct eth_hdr *) m->data;
to 
struct eth_hdr * eth = (struct eth_hdr *) ( (char *)m->buf _addr + m->data_offset);

We have some code that actually processes 4-8 packets in parallel (parse + hash), with a pretty high IPC. What we've done here is essentially replaced is a simple load, with  a load, load, add sequence in front of it. There is no real way to do these computations in parallel for multiple packets - it has to be done one or two at a time. What suffers is the IPC of the overall function that does the parse/hash quite significantly. It's those functions that I worry about more than the driver.  I haven't yet been able to come up with a mitigation for this yet. 

Neil, 

The last time we looked at this change - and it's been a while ago, the negative effect on the upper level functions built on this was on the order of about 15-20%. It's probably will get worse once we tune the code even more.  Hope the above explanation gives you a flavour of the problem this will introduce. 

Regards, 
-Venky




-----Original Message-----
From: Olivier MATZ [mailto:olivier.matz at 6wind.com] 
Sent: Monday, May 12, 2014 8:07 AM
To: Neil Horman; Venkatesan, Venky
Cc: Thomas Monjalon; dev at dpdk.org
Subject: Re: [dpdk-dev] [PATCH RFC 06/11] mbuf: replace data pointer by an offset

Hi Venky,

On 05/12/2014 04:41 PM, Neil Horman wrote:
>> This is a hugely problematic change, and has a pretty large 
>> performance impact (because the dependency to compute and access). We 
>> debated this for a long time during the early days of DPDK and 
>> decided against it. This is also a repeated sequence - the driver 
>> will do it twice (Rx + Tx) and the next level stack will do it twice 
>> (Rx + Tx) ...
>>
>> My vote is to reject this change particular change to the mbuf.
>>
>> Regards,
>> -Venky
>>
> Do you have perforamance numbers to compare throughput with and 
> without this change?  I always feel suspcious when I see the spectre 
> of performane used to support or deny a change without supporting reasoning or metrics.

I agree with Neil. My feeling is that it won't impact performance, and it is correlated with the forwarding tests I've done with this patch.

I don't really understand what would cost more by storing the offset instead of the virtual address. I agree that each time the stack will access to the begining of the mbuf, there will be an arithmetic operation, but it is compensated by other operations that will be
accelerated:

- When receiving a packet, the driver will do:

     m->data_off = RTE_PKTMBUF_HEADROOM;

   instead of:

     m->data = (char*) rxm->buf_addr + RTE_PKTMBUF_HEADROOM;

- Each time the stack will prepend data, it has to check if the headroom
   is large enough to do the operation. This will be faster as data_off
   is the headroom.

- When transmitting a packet, the driver will get the physical address:

     phys_addr = m->buf_physaddr + m->data_off

   instead of:

     phys_addr = (m->buf_physaddr +  \
         ((char *)m->data - (char *)m->buf_addr)))

Moreover, these operations look negligible to me (few cycles) compared to the large amount of arithmetic operations and tests done in the driver.

Regards,
Olivier