binary protocol notes from the facebook hackathon

‏Alex Stapleton alexs at
Wed Jul 11 11:04:50 UTC 2007

On 11 Jul 2007, at 11:56, Steve Grimm wrote:

> On 7/11/07 12:45 AM, "marc at" <marc at> wrote:
>> While network byte ordering (Big Endian) is traditionally the 'right'
>> thing to do (or the default thing to do), in most cases it's a minor
>> performance hit due to constant swapping.  Since we're implementing a
>> binary protocol specifically to avoid/minimize minor performance hits
>> and since this is a brand new protocol I would recommend to keep all
>> values as Little Endian because:
> Defining a protocol that uses something other than network byte  
> order gives
> me the willies. But I'm not sure I can argue with your reasoning on  
> anything
> more than aesthetic grounds; swapping bytes around unnecessarily is  
> exactly
> the sort of thing we're trying to avoid.
> I wonder just what the overhead of byte swapping will be, though,  
> compared
> to all the other work that has to happen to handle a request. It  
> may be a
> small enough expense to disappear into the statistical noise. If  
> we'd be
> saving .001% of the CPU time by avoiding swapping, I'd say forget  
> it, not
> worth the confusion.  I've never actually measured the cost of  
> using network
> byte order in any of my network apps, though, so I honestly have no  
> idea.
> If it's a large overhead, maybe we can make endianness depend on  
> the magic
> number at the start of the command. It is annoying to support both,  
> granted,
> but that would potentially allow a client to be configured to send  
> requests
> and receive responses in whatever format is most efficient for a  
> particular
> server. (Or vice versa, I suppose.) Or would the overhead of having to
> support both styles eat all the cycles you'd save by not swapping  
> bytes? I'm
> not sure.
> I suspect that the Sun folks would appreciate not having to flip  
> bits back
> to network byte order on their SPARC boxes, but I agree with you  
> that Intel
> and AMD totally dominate memcached's target audience at the moment.
> Supporting both orders from the start would insulate us from  
> changes in that
> market reality, though maybe that's far enough off that it's better  
> to just
> rev the protocol if/when the need arises.
> -Steve

Make the message IDs different for network byte ordered requests /  
responses than the host order ones. That way they can take a separate  
code path with no additional overhead and the client can request data  
in whatever byte ordering it wants. It halves the number of message  
types we can support, do we really even need 127 message types? I  
hope not :)

More information about the memcached mailing list