Why no amd-64 nano-BSD version?



  • With Atom-D510 based systems like the Lanner FW-7535 which can take a little CF card to boot from, it would make sense to have such a configuration, it would seem to me.
    Is there a reason why that's not supplied?


  • Rebel Alliance Developer Netgate

    Any reason why you really need to run amd64 on a little Lanner box? Does it really have more than 4GB of RAM?

    There is an open ticket for it, iirc, but there is no real compelling reason that amd64 is required in most scenarios except to access large amounts of RAM.



  • @jimp:

    Any reason why you really need to run amd64 on a little Lanner box? Does it really have more than 4GB of RAM?

    There is an open ticket for it, iirc, but there is no real compelling reason that amd64 is required in most scenarios except to access large amounts of RAM.

    I was under the impression that the extra register available in the amd64 programming model make compilers vastly more efficient. There was a developer talk regarding Mac OS X and they showed how the same code was running 30-50% faster when simply being recompiled for amd64, because the compiler could use register based value passing rather than doing the stack model.

    They said in all, but a few exceptional cases where lots of byte processing is done, the 64-bit CPU instruction set makes for considerably more efficient code.

    Of course, that could have been specific to the compilers Apple uses and FreeBSD might be a slightly different situation, or the kind of processing that pfSense does might be one of the exceptions.

    But if their observations apply also to this platform, then with even low-end Atom based boxes now supporting the 64-bit instruction set, it would seem to be a worthwhile cause, regardless of the amount of memory present.


  • Rebel Alliance Developer Netgate

    Unfortunately we don't have any actual real-world numbers to support or refute that.

    Nobody I am aware of has done benchmarks between the two versions on identical hardware.



  • @jimp:

    Unfortunately we don't have any actual real-world numbers to support or refute that.

    Nobody I am aware of has done benchmarks between the two versions on identical hardware.

    Might be a worthwhile setup to come up with some benchmark configuration e.g. looping various traffic out one port and in another, so the testing could be done on a single box.

    That would allow to create some sort of index to allow assessment of relative performance among various devices used for running pfSense.

    Right now it's not trivial to guess how big/expensive a box one should consider without being underpowered or wasting money and electricity.

    Once a benchmark config is defined, it would be trivial to test boxes running on or the other version of the OS and compare max. throughput or system load under a fixed traffic load.


  • Rebel Alliance Developer Netgate

    You're assuming we have some giant network lab with tons of hardware for running such things… Such a thing doesn't exist. :-)

    We rely mostly on what users in the community report, or what we can observe in the field. Generally people haven't had a need to try both amd64 and i386. We can test a few things, but many benchmarks are meaningless when used in place of real world traffic scenarios.



  • @jimp:

    You're assuming we have some giant network lab with tons of hardware for running such things… Such a thing doesn't exist. :-)

    We rely mostly on what users in the community report, or what we can observe in the field. Generally people haven't had a need to try both amd64 and i386. We can test a few things, but many benchmarks are meaningless when used in place of real world traffic scenarios.

    No, what I was actually thinking about was this: most pfSense configurations will have at least two NICs.
    So there could be some test configuration that tests throughput/system load when connecting the two NICs of the same system with a cross-over cable.
    That way one could e.g. measure CPU load when saturating the interface with traffic of various kinds, or something like that.


  • Rebel Alliance Developer Netgate

    That tells you nothing.

    pfSense is optimized to be a router, not an endpoint. So unless the traffic is sent to/from devices on interfaces, the benchmark would be pretty meaningless. It doesn't tell you how much traffic the box can pass because you also have the overhead of whatever utility is generating the traffic.

    You really need three systems - the router, a WAN system, and a LAN system, and you can run things like iperf between the WAN/LAN clients to test things and get better numbers.


Log in to reply