[Soekris] Soekris net6501-70: CPU Speed / Network throughput slow?
Kyle Brantley
kyle at averageurl.com
Wed Nov 9 16:08:01 UTC 2011
On 11/9/2011 8:10 AM, Lonnie Abelbeck wrote:
> Hi Hermann,
>
> I am getting better results than you are.
>
> I have a the 1MB/1GHz net6501-50, comBIOS 1.40h, and using it as a router...
>
> [ iperf -s ] --- [eth0/WAN # net6501-50 # eth2/LAN] --- [ iperf -c ... ]
>
> I get 515 Mbits/sec with the default iperf settings. 'top' is non-responsive during the test on the console.
>
> I'm using linux kernel 2.6.35 with the latest e1000e driver from sourceforge, no-SMP and iptables (IPv4 NAT/IPv6) firewall active.
> --
> # cat /proc/interrupts
> CPU0
> 0: 243 IO-APIC-edge timer
> 2: 0 XT-PIC-XT cascade
> 4: 416 IO-APIC-edge serial
> 14: 0 IO-APIC-edge ide0
> 15: 0 XT-PIC-XT ide1
> 16: 0 IO-APIC-fasteoi ehci_hcd:usb2, ohci_hcd:usb6, ohci_hcd:usb7, ohci_hcd:usb8
> 19: 0 IO-APIC-fasteoi ehci_hcd:usb1, ohci_hcd:usb3, ohci_hcd:usb4, ohci_hcd:usb5
> 40: 1281 PCI-MSI-edge ahci
> 41: 150927 PCI-MSI-edge eth0-Q0
> 42: 3 PCI-MSI-edge eth0
> 45: 291571 PCI-MSI-edge eth2-Q0
> 46: 9 PCI-MSI-edge eth2
> 47: 30384 PCI-MSI-edge eth3-Q0
> 48: 2 PCI-MSI-edge eth3
> NMI: 0 Non-maskable interrupts
> LOC: 60801192 Local timer interrupts
> SPU: 0 Spurious interrupts
> PMI: 0 Performance monitoring interrupts
> PND: 0 Performance pending work
> TRM: 0 Thermal event interrupts
> THR: 0 Threshold APIC interrupts
> MCE: 0 Machine check exceptions
> MCP: 203 Machine check polls
> ERR: 0
> MIS: 0
> --
>
> Extrapolating, your net6501-70 should get near 1Gbit wire-speed routing performance.
>
> Lonnie
>
>
>
> On Nov 9, 2011, at 6:29 AM, Hermann Himmelbauer wrote:
>
>> Hi,
>> I am just testing the new Soekris net6501-70 with Debian 6.0.3 (Kernel
>> 2.6.32-5). I had a net5501-70 before and upgraded to be able to perform
>> gigabit routing.
>>
>> During testing, I also measured the network throughput with "iperf" and
>> experienced the following:
>>
>> - When measuring the throughput between my desktop PC and the net6501, I
>> get approx. 290MBits/sec. It is interesting that "iperf" has 100% CPU
>> usage during this test (measured via top). On the desktop PC, iperf
>> consumes approx. 10%.
>> - When using the net6501 as a router, the throughput is also at
>> 290Mbits/sec, but "top" is stuck during the test on the net6501, the
>> machine is completely stuck.
>>
>> On the net5501-70, throughput is at ~ 95Mbits/sec (as it has no gigabit
>> ethernet) and the CPU loads are similar, but approx 1/3rd of the above
>> (30% with iperf). Measuring the CPU load during routing seems not to be
>> possible via top.
>>
>> Suspecting the CPU, I decided to do some CPU testing, using nbench
>> (nbench uses one core only!). Below are the results / comparisons for my
>> desktop PC, the net6501-70 and net5501-70:
>>
>> ===========================LINUX DATA BELOW==========================
>> CPU : Dual GenuineIntel Intel(R) Core(TM)2 CPU
>> 6300 @ 1.86GHz 1862MHz
>> L2 Cache : 2048 KB
>> OS : Linux 2.6.37.6-0.5-desktop
>> C compiler : gcc version 4.3.2 (Debian 4.3.2-1.1)
>> libc : libc-2.7.so
>> MEMORY INDEX : 16.325
>> INTEGER INDEX : 15.627
>> FLOATING-POINT INDEX: 24.542
>> Baseline (LINUX) : AMD K6/233*, 512 KB L2-cache, gcc 2.7.2.3, libc-5.4.38
>>
>> =========================LINUX DATA BELOW============================
>> CPU : GenuineIntel Genuine Intel(R) CPU @ 1.60GHz
>> 1600MHz
>> L2 Cache : 512 KB
>> OS : Linux 2.6.32-5-686
>> C compiler : gcc version 4.4.5 (Debian 4.4.5-8)
>> libc : libc-2.11.2.so
>> MEMORY INDEX : 2.928
>> INTEGER INDEX : 2.873
>> FLOATING-POINT INDEX: 2.631
>> Baseline (LINUX) : AMD K6/233*, 512 KB L2-cache, gcc 2.7.2.3, libc-5.4.38
>>
>> ==========================LINUX DATA BELOW==========================
>> CPU : AuthenticAMD Geode(TM) Integrated Processor by AMD
>> PCS 500MHz
>> L2 Cache : 128 KB
>> OS : Linux 2.6.33.1-geode
>> C compiler : gcc version 4.4.5 (Debian 4.4.5-8)
>> libc : libc-2.11.2.so
>> MEMORY INDEX : 1.919
>> INTEGER INDEX : 1.887
>> FLOATING-POINT INDEX: 1.595
>>
>>
>> I find it interesting that the difference between the Intel Core and
>> Atom CPUs is that high. I expected the Atom to have e.g. 50% of the
>> speed of an Intel Core, but something like 10% to 20% is surprising.
>> Although the clock speed of the AMD geode processor is far lower than
>> the Atom and the processor generation is way behind, it still has
>> approx. 65% of its speed.
>>
>> On the other hand, when looking at /proc/cpuinfo, the Core2 has 3878
>> bogomips, while the Atom has 3200 bogomips and the AMD Geode has 1000,
>> which sounds a lot more appropriate.
>>
>> So I wonder if there's something wrong with my setup? Perhaps the kernel
>> version I use has problems with the net6501? Or maybe there is something
>> wrong with the bios? Maybe the CPU gets too hot and is throttling? The
>> surface of the heatsink has 50°C, which is not cool, but I think still
>> not too hot. And I tried cooling it down to ~ 38°C for the test, which
>> made no difference.
>>
>> Any ideas/recommendations? Can anyone achieve a network throughput of
>> something near 1Gbit with the net6501-70?
>>
>> Best Regards,
>> Hermann Himmelbauer
>> _______________________________________________
>> Soekris-tech mailing list
>> Soekris-tech at lists.soekris.com
>> http://lists.soekris.com/mailman/listinfo/soekris-tech
>>
>>
> _______________________________________________
> Soekris-tech mailing list
> Soekris-tech at lists.soekris.com
> http://lists.soekris.com/mailman/listinfo/soekris-tech
A quick test against a device on the same switch as my 6501-70, ran for
two minutes:
[ ID] Interval Transfer Bandwidth
[ 3] 0.0-120.0 sec 11.1 GBytes 797 Mbits/sec
Linux localhost.localdomain 2.6.38.6-26.rc1.fc15.x86_64 #1 SMP Mon May 9
20:45:15 UTC 2011 x86_64 x86_64 x86_64 GNU/Linux
Not quite full wire speed, but acceptable. It also consumed the 'entire'
CPU (as per top, 100% dedicated to software interrupts), leaving the HT
core mostly available.
--Kyle
More information about the Soekris-tech
mailing list