Ubuntu: A follow-up on 32-bit powerpc architecture

Posted Mar 18, 2017 5:29 UTC (Sat) by allesfresser (guest, #216)
In reply to: Ubuntu: A follow-up on 32-bit powerpc architecture by pizza
Parent article: Ubuntu: A follow-up on 32-bit powerpc architecture

You would think so, but perhaps except for the Firewire interface it's pretty poky overall. I find myself getting frustrated with its slowness far more than the Pi3. I only keep using it out of a perverse sense of nostalgia, I suppose. :p

Ubuntu: A follow-up on 32-bit powerpc architecture

Posted Mar 18, 2017 5:42 UTC (Sat) by allesfresser (guest, #216) [Link] (10 responses)

Benchmarks for G4 Mini: http://browser.primatelabs.com/geekbench2/2363348 and Raspi 3: http://browser.primatelabs.com/geekbench2/2596704 Just for idle comparison fun.

Ubuntu: A follow-up on 32-bit powerpc architecture

Posted Mar 18, 2017 11:57 UTC (Sat) by pizza (subscriber, #46) [Link] (8 responses)

That's why I was careful to say I/O performance. :)

For a while I was supporting a series of boards built using IBM/AMCC PPC405ep SoCs. They sported dual onboard 100Mbps ethernet ports wired directly into the CPU's DMA matrix, and one or two miniPCI slots. Even at a languid 266MHz clock speed, the system could drive both ethernet interfaces at full-wire speed and still have more than enough oomph left over to max out at least one 802.11a/g device with the Linux bridging/routing/iptabling/etc everything.

Meanwhile, even the modern RPi3 is stuck sipping all of its I/O through a single USB2 interface that can't even max out a single 100Mbps ethernet device. (Yeah, I realize it also has SD/SDIO, but that performance is barely better than what a USB stick can achieve on the same system.)

Sure, it's hard to argue with the RPi's price and excellent support/etc, but I'm far from the only one who wished Broadcom had at least slapped a second (even crummy) USB interface (or improved the one's performance) on the later generations of the SoC instead of just replacing the CPU cores with something less anaemic.

(Next week I'm going to be doing some bus-level analysis of some communication failures involving the RPi and a certain family of printers. While I'm at it I may see why certain USB wifi devices also keel over under load too..)

Ubuntu: A follow-up on 32-bit powerpc architecture

Posted Mar 18, 2017 13:05 UTC (Sat) by allesfresser (guest, #216) [Link] (6 responses)

Oh, I absolutely agree about wishing the Pi had better I/O channels (SATA, GigE, etc.) but for the low-cost education mission they designed it for, they've done a terrific job.

Ubuntu: A follow-up on 32-bit powerpc architecture

Posted Mar 18, 2017 18:47 UTC (Sat) by jch (guest, #51929) [Link] (5 responses)

There are a number of boards that have better I/O than the raspberry Pi, and run with few or no binary blobs. They usually don't have such an extensive and friendly community as the Pi, though, and they tend to have slower CPUs than the Pi 3.

If you want a kernel that's being actively maintained and a mostly free userspace (with the exception of some fairly exotic devel tools), the various BeagleBone variants are pretty nice — a single 100Gbit Ethernet phy on the base model, but directly attached to the SoC over GMII (the wireless models require a binary blob for the WiFi firmware). If you need gigabit Ethernet, Olimex are building some rather nice boards (e.g. the A20-Lime2). If you need more extensive networking but can manage without GPIOs, nothing can beat the mass-market routers that happen to be well supported by OpenWRT (the venerable WNDR3800 still being my favourite, but more recent models have vastly superior CPUs).

Ubuntu: A follow-up on 32-bit powerpc architecture

Posted Mar 18, 2017 22:54 UTC (Sat) by pizza (subscriber, #46) [Link]

My griping with the RPi has more to do with the fact that Broadcom has now released two significant updates to the SoC powering the thing, significant boosting the CPU capabilities and performance each time, but with essentially no change to rest of the SoC.

Sure, anything major would have resulted in something that's no longer [mostly] pin-compatible, but would it have really been that hard to replace the BM2385's afterthought-of-a-USB controller with something a bit less horrid? (...Especially given that Broadcom already has other USB controllers that are much, much more capable...)

OpenWRT device recommendation?

Posted Mar 20, 2017 21:03 UTC (Mon) by moxfyre (guest, #13847) [Link] (3 responses)

Do you have a particular OpenWRT-compatible device that you'd recommend for general hacking around with embedded Linux and networking as of 2017? Ideally something that is fast enough to keep up with 100 Mbps of aes128 (even better if hardware accelerated :-D).

I have an OpenWRT device but it's ancient and I'm having trouble figuring out what to replace it with.

OpenWRT device recommendation?

Posted Mar 22, 2017 1:18 UTC (Wed) by flussence (guest, #85566) [Link] (2 responses)

PCEngines does embedded boards with hardware AES128 (Geodes and newer AMD chips). Enough to saturate the Ethernet/IDE interfaces, it's probably more bottlenecked by the RAM.

But given it's 2017, you should probably be considering updating your benchmarks for Chacha20, which is more than fast enough by your criteria (`openssl speed chacha20-poly1305` reports 35MB/s on a 1st-gen Atom). It's been the default cipher in OpenSSH for two years and a large chunk of the web uses it too — you're likely using it more than AES at this point.

OpenWRT device recommendation?

Posted Mar 23, 2017 13:38 UTC (Thu) by mstone_ (subscriber, #66309) [Link] (1 responses)

which is sad, since AES-NI assisted AES-GCM is an order of magnitude faster on modern CPUs

OpenWRT device recommendation?

Posted Mar 24, 2017 0:15 UTC (Fri) by flussence (guest, #85566) [Link]

Faster, but given the ongoing tsunami of revelations about TLA agencies, is it such a good idea to outsource all security to a magic black box on the CPU?

Ubuntu: A follow-up on 32-bit powerpc architecture

Posted Mar 23, 2017 20:19 UTC (Thu) by derobert (subscriber, #89569) [Link]

At least according to my tests, that's not true. A Raspberry Pi 3 can saturate its 100mbps NIC.

Using iperf3, I tested my RPi3 to a PC on the LAN (the LAN and the PC are gigabit):

- Up or down, alone: 94.5mbps TCP / 95.8 (UDP)
- Simultaneous (two iperf3s running, one with -R, two iperf3 servers running): 81.1/92 mbps (TCP) / 94.6/95.6 mbps (UDP)

That's really about as good as you can expect on a 100mbps Ethernet interface.

Ubuntu: A follow-up on 32-bit powerpc architecture

Posted Mar 18, 2017 13:29 UTC (Sat) by ianmcc (subscriber, #88379) [Link]

Interesting that some of the floating point benchmarks for the RPi are worse than the mac mini. It is really only he multi-cores of the RPi that save it. And the sharpen / blur benchmarks. What is different about them? Are they using the GPU?