3 ms·

An alternative, if you want to truly get it manufactured yourself (the schematics and materials list are all open source): https://www.turris.cz/en/hardware ht

An alternative, if you want to truly get it manufactured yourself (the schematics and materials list are all open source): https://www.turris.cz/en/hardware https://www.turris.cz/en/hardware This device is a lot more powerful than the PC Engines APU. As nice as the APU is, its CPU cannot support 1Gbps throughput over the NICs.

wtallis 12y ago
Got any references for the relative performance of that over the APU board? They're not using the AMD Geode processors anymore. What you linked is a dual-core 1.2GHz 32-bit PPC built on a 45nm process. The APU used on the board I linked to is a dual-core 1GHz x86_64 built on a 40nm process. There's no way the PPC core is dramatically faster than the x86 core that was shown to have better IPC than Intel's Atom. The PPC SoC might be able to offload a lot of network stuff to fixed-function hardware, but one of the primary lessons learned from the CeroWRT project is that you can't trust any of the offload stuff to be well-behaved with respect to things like queue length management.
- samcrawford 12y ago
  Yes. The APU maxes out around ~670Mbit/s using iperf [1]. I confirmed as much in correspondence with PC Engines too. Personally I _suspect_ it should be able to go much faster, but haven't got one to investigate yet. The 1.2GHz PPC based router is hitting ~940Mbit/s (the most you can expect from a GE interface with protocol overheads) using iperf. I don't have a reference for that, but I've got one of the devices here and have performed that test myself. 1. http://planet.ipfire.org/post/pc-engines-apu1c-a-review http://planet.ipfire.org/post/pc-engines-apu1c-a-review
  - dtaht 12y ago
    Thank you for the performance statistics. We have done a lot of work on other hardware platforms to get them up to being able to saturate gigE or higher, with tuning and things like "Byte Queue Limits" support. The problem with the ppc thing is that most likely it has bufferbloat in a blob we can't fix. (news to the contrary welcomed). And hearing that the APU peaks at 670Mbit is disappointing, but perhaps someone will get some time to profile it and figure out why it is so bad - certainly we get gigE easily with intel based hardware.

wtallis 12y ago

Got any references for the relative performance of that over the APU board? They're not using the AMD Geode processors anymore. What you linked is a dual-core 1.2GHz 32-bit PPC built on a 45nm process. The APU used on the board I linked to is a dual-core 1GHz x86_64 built on a 40nm process. There's no way the PPC core is dramatically faster than the x86 core that was shown to have better IPC than Intel's Atom. The PPC SoC might be able to offload a lot of network stuff to fixed-function hardware, but one of the primary lessons learned from the CeroWRT project is that you can't trust any of the offload stuff to be well-behaved with respect to things like queue length management.

samcrawford 12y ago
Yes. The APU maxes out around ~670Mbit/s using iperf [1]. I confirmed as much in correspondence with PC Engines too. Personally I _suspect_ it should be able to go much faster, but haven't got one to investigate yet. The 1.2GHz PPC based router is hitting ~940Mbit/s (the most you can expect from a GE interface with protocol overheads) using iperf. I don't have a reference for that, but I've got one of the devices here and have performed that test myself. 1. http://planet.ipfire.org/post/pc-engines-apu1c-a-review http://planet.ipfire.org/post/pc-engines-apu1c-a-review
- dtaht 12y ago
  Thank you for the performance statistics. We have done a lot of work on other hardware platforms to get them up to being able to saturate gigE or higher, with tuning and things like "Byte Queue Limits" support. The problem with the ppc thing is that most likely it has bufferbloat in a blob we can't fix. (news to the contrary welcomed). And hearing that the APU peaks at 670Mbit is disappointing, but perhaps someone will get some time to profile it and figure out why it is so bad - certainly we get gigE easily with intel based hardware.