Intel Avoton C2750 Benchmarks - Supermicro A1SAi-2750F

hapi

New Member
Sep 28, 2013
5
0
0
sorry, my fault with 82574L.

interrupt handling is very difficult when 100kpps. Not only used one core for handling interrupts but it is spread across multiple cores. More power for the router. Tested. Processing packets inside the kernel does not interest me. I can not influence. Current Ethernet with one queue (1tx/rx) Kill reliably router around 10kpps. In exchange for i350 (4tx/4rx) grow extremely high maximum load. Also tested. Packet processing in the kernel, it is not so difficult, especially when each core handles a quarter rx of the flow through the delivery of the packet to a particular core.

dual Ethernet with 1 queue rx and 1 tx

# CPU LOAD IRQ DISK
0 CPU0 0% 0% 0%
1 CPU1 0% 0% 0%
2 CPU2 0% 0% 0%
3 CPU3 0% 0% 0%
4 cpu4 28% 28% 0%
5 cpu5 0% 0% 0%
6 cpu6 3% 3% 0%
7 cpu7 21% 21% 0%


dual Ethernet with 4 rx queue and 4 tx

# CPU LOAD IRQ DISK
0 CPU0 8% 8% 0%
1 CPU1 5% 5% 0%
2 CPU2 5% 5% 0%
3 CPU3 4% 4% 0%
4 cpu4 4% 4% 0%
5 cpu5 4% 4% 0%
6 cpu6 9% 9% 0%
7 cpu7 6% 6% 0%

both the Xeon E3.

Anyway, thank you all for your reply.
 

Marco

New Member
Sep 23, 2013
17
0
1
No offense man, and sorry for being rough, but nobody needs your (meaningless and useless [1]) data to discover something obvious and well known since many years (see Linux NAPI, hw interrupt moderation, polling based software routers, etc.), especially to people like me that design routers. What you seem to fail to understand, though, is that the ethernet controller chooses how to enqueue packets into different descriptor rings looking at the traffic. To make a long story short, it doesn't really matters what policy/filters you select, whether priority based, hash based, vlan based, mac based, port based and so on, the important thing is that traffic must come with enough "differences" to be spread among different queues. This usually means lots of different sources with different traffic characteristics. And even then, it's frequent to see some cores more loaded than others.
That said I have no idea about your use case, but I just said (once again) that despite the fact that this Avoton motherboard has plenty of cores it might perform poorly in some common applications (e.g. fileserver), at least it might not perform as good as a low power Xeon. It would be nice to see some well done comparison.


[1] what about the packet size, the traffic characteristics, the testbed? And so on...
 
Last edited:

hapi

New Member
Sep 28, 2013
5
0
0
No offense man, and sorry for being rough, but nobody needs your (meaningless and useless [1]) data to discover something obvious and well known since many years (see Linux NAPI, hw interrupt moderation, polling based software routers, etc.), especially to people like me that design routers. What you seem to fail to understand, though, is that the ethernet controller chooses how to enqueue packets into different descriptor rings looking at the traffic. To make a long story short, it doesn't really matters what policy/filters you select, whether priority based, hash based, vlan based, mac based, port based and so on, the important thing is that traffic must come with enough "differences" to be spread among different queues. This usually means lots of different sources with different traffic characteristics. And even then, it's frequent to see some cores more loaded than others.
That said I have no idea about your use case, but I just said (once again) that despite the fact that this Avoton motherboard has plenty of cores it might perform poorly in some common applications (e.g. fileserver), at least it might not perform as good as a low power Xeon. It would be nice to see some well done comparison.


[1] what about the packet size, the traffic characteristics, the testbed? And so on...

not counting the load caused by filters, nat, mangle, etc.. Counting only the strains generated IRQ. More on ethernet IRQ They help distribute load among multiple cores. We have the experience that Ethernet is capable of doing this, can learners to enhance the throughput of routers and are more resistant to DDoS through them.
 

gonzopancho

New Member
Nov 17, 2013
4
0
1
igb driver, awesome, and part of the i350 family it seems so all the VM feature goodness.

The linux kernel for these should be mainstream pretty quick depending how much your distro loves to do updates.

Since its intel driver based the BSDs might get it this year in stable releases instead of next year (or never) for marvell, which means pfsense will have it around 2015 instead of 2020. (only slightly sarcastic)
You can be as sarcastic as you like, but we're actively working on exactly that board (at Netgate, but since I own over half of the company behind pfSense... well, draw your own conclusion.)
 

Patrick

Administrator
Staff member
Dec 21, 2010
11,962
4,921
113
You can be as sarcastic as you like, but we're actively working on exactly that board (at Netgate, but since I own over half of the company behind pfSense... well, draw your own conclusion.)
It is a great board. Still need to pick up the Rangeley version.

pfsense rocks btw!
 

gonzopancho

New Member
Nov 17, 2013
4
0
1
I'm not holding my breath with pfsense support considering it still doesn't support 802.11N (couple exceptions). If Linux support for the NICs is tricky, BSD is going to be impossible for awhile. Of course, the way around this is virtualization through ESXI.
If only you knew what you were talking about.. (sigh)

pfSense started to lag because the then lone leader isn't a programmer. The situation changed last year about this time. Now that pfSense 2.1 (based on FreeBSD 8.3) is out, a port to FreeBSD 10 is underway (and working internally). FreeBSD 10 has a lot of work to improve the situation WRT 802.11n.
 

nitrobass24

Moderator
Dec 26, 2010
1,083
127
63
TX
@gonzo - Any chance of us seeing a Linux port of PFSense? I really want to like PFSense, but my entire setup is based on Xen, which unfortunately does not support BSD and network intense applications seem to really suffer since there are no VM tools for BSD.

I'm pretty sure I already know the answer, but it doesn't hurt to ask? :)

@Patrick - When are you getting a Rangley board? Im curious if you can use non-ecc RAM.
 

Patrick

Administrator
Staff member
Dec 21, 2010
11,962
4,921
113
@gonzo - Any chance of us seeing a Linux port of PFSense? I really want to like PFSense, but my entire setup is based on Xen, which unfortunately does not support BSD and network intense applications seem to really suffer since there are no VM tools for BSD.

I'm pretty sure I already know the answer, but it doesn't hurt to ask? :)

@Patrick - When are you getting a Rangley board? Im curious if you can use non-ecc RAM.
Well.. was supposed to be two weeks ago. Hoping this week. Scheduling! My guess is that Rangeley will be very similar to Avoton in terms of memory support. Would strongly suggest using ECC though just because non-ECC has been nothing but trouble for the three Avoton units on hand.
 

randomdood

New Member
Nov 19, 2013
2
0
0
Would such BMC chip allow Direct3D Acceleration?
(DXDIAG Tab Monitor (n))

*edit*
Would such BMC chip allow passthrough in vmware / xen or others?
 
Last edited:

hjfr

Member
Nov 21, 2013
76
20
8
France
Hi Patrick, could you post a "cat /proc/cpuinfo" from your c2750 test platform ?
It's to show cpu capabilities (flags) because intel ARK not provide many informations ...
Thanks in advance.
 

Jeggs101

Well-Known Member
Dec 29, 2010
1,489
228
63
Hi Patrick, could you post a "cat /proc/cpuinfo" from your c2750 test platform ?
It's to show cpu capabilities (flags) because intel ARK not provide many informations ...
Thanks in advance.
I would be interested in this too.
 

Patrick

Administrator
Staff member
Dec 21, 2010
11,962
4,921
113
Finally was able to pick up the Rangeley Supermicro A1SRi-2758F. Will see if I can get that done this evening.
 

hjfr

Member
Nov 21, 2013
76
20
8
France
Thanks. Very interesting.

The difference between c2750 and c2758 are no turbo and QuickAssist for c2758 ?
 

Patrick

Administrator
Staff member
Dec 21, 2010
11,962
4,921
113
Thanks. Very interesting.

The difference between c2750 and c2758 are no turbo and QuickAssist for c2758 ?
Those are the key differences from what I have seen thus far.
 

rthorntn

Member
Apr 28, 2011
33
4
8
Hello Everyone,

This beauty just arrived, once the 32GB of ECC RAM arrives I will be putting ESXi 5.5 on it with the upgraded VIB (thanks for the info on that Patrick)!

It's going in a tiny case because I need it to be very portable, the Mini-Box M350.

It will have two SSD's installed so no space for fans on the top brackets and to keep the noise down I don't want to use the front 40mm opening, so that really leaves somehow securely fixing a low speed 60mm fan to the top of the supplied heatsink (it's 60x55mm), does anyone have any ideas?

Thanks!

Richard
 

Patrick

Administrator
Staff member
Dec 21, 2010
11,962
4,921
113
Richard, I sent an e-mail to see if I can find a part number for the fan that came with my Avoton board. Rumor has it that the fan is used by their engineers.