Supermicro Intel® Atom™ Processor C2758

LostInIgnorance

I am looking at redoing my dual-core sustem that I seem to be maxing out at times. I am running snort, squid, lightsquid, havp antivirus and nmap packages. I am looking at possibly purchasing this with a couple of SSD (mirrored) drives.

http://www.supermicro.com.tw/products/system/1u/5018/sys-5018a-ftn4.cfm

Anyone think it'll be under-powered (peaking out) with 2 100mb (load balanced) connections incoming and the packages above?

asterix

I haven't tested the C2758s yet but my typical recommendation has always been to go for an i3. There is not much of power saving difference or in other words you won't be a $100 richer after running an atom or other CPU slower than the i3. The processor itself slows down when not required and has the added juice to push snort, dansguardian with clamd, squid and other packages all at the same time.

jasonlitka

@LostInIgnorance:

I am looking at redoing my dual-core sustem that I seem to be maxing out at times. I am running snort, squid, lightsquid, havp antivirus and nmap packages. I am looking at possibly purchasing this with a couple of SSD (mirrored) drives.

http://www.supermicro.com.tw/products/system/1u/5018/sys-5018a-ftn4.cfm

Anyone think it'll be under-powered (peaking out) with 2 100mb (load balanced) connections incoming and the packages above?

Search for my posts on the subject. That chip is fast. Very fast. 2x100 is nothing.

stephenw10

There are few people running/testing that box (including ESF I believe) so it should be possible to get some good numbers.
The 8 core CPU will be much better utilised when 2.2 is released with the new SMP PF.
Those new Atoms are lightyears away from the old Atoms but still 200Mbps with Snort Squid and HAVP… :-\ I wouldn't like to guess.

What are you updating from? Is it running out of cpu cycles?

Edit: Typed too slow! Jason has this box so see above. ::)

Steve

jdillard

It is also sold in the pfSense store, bundled with a year of support from the pfsense team.

http://store.pfsense.org/c2758/

jasonlitka

@jdillard:

It is also sold in the pfSense store, bundled with a year of support from the pfsense team.

http://store.pfsense.org/c2758/

You guys should retest that system. My tests show it's capable of a LOT more, even without AES-NI. I don't know about QuickAssist, I haven't explicitly turned it on but I guess it could be helping in the background.

cmb

@Jason:

You guys should retest that system. My tests show it's capable of a LOT more, even without AES-NI. I don't know about QuickAssist, I haven't explicitly turned it on but I guess it could be helping in the background.

QuickAssist isn't supported (yet), so at the moment it won't do anything.

We will be doing more testing on that system, including with LACP as I suspect it can probably do 2 Gb wire speed TCP (numbers there now are limited at 1 Gb wire speed).

What kind of numbers were you seeing?

jasonlitka

@cmb:

@Jason:

You guys should retest that system. My tests show it's capable of a LOT more, even without AES-NI. I don't know about QuickAssist, I haven't explicitly turned it on but I guess it could be helping in the background.

QuickAssist isn't supported (yet), so at the moment it won't do anything.

We will be doing more testing on that system, including with LACP as I suspect it can probably do 2 Gb wire speed TCP (numbers there now are limited at 1 Gb wire speed).

What kind of numbers were you seeing?

The original numbers I posted in the hardware section were only after a few days of light testing on a slower connection but I've since upgraded by FiOS (for the second time this year, don't tell my wife) and I think I've got a better idea of what it can do.

I've got an OpenVPN tunnel setup between my home and a Linode server in Newark using BF-CBC (this is not a privacy thing, it's a "my NetFlix works better thing"). Ping between locations is ~14ms. Running full out on my connection I see 135Mbit/s on speedtest.net to a server in NYC. The CPU usage on my box for openvpn was ~34% of a core. Assuming that CPU usage is linear with throughput, my math says this will get close to 400Mbit/s per OpenVPN connection without any hardware acceleration. Even if it's not linear, I'm already doing better than you guys say is the max and I'm doing it with 1/3 of a core.

I've also done some tests with Snort. I know this one is a bit harder to quantify since the performance varies greatly based on what rules you've picked and what mode you've got it in, but in AC mode with the VRT Balanced rules, plus a dozen or so ET rulesets, I get ~55% CPU at 135Mbit/s which would put us at ~250Mbit/s. My earlier tests were closer to double that but I added a BUNCH of extra rules recently.

Are you sure your box wasn't overheating or wasn't throttled back? I've noticed that in "adaptive" it's really hesitant to speed back up. I've been using "hiadaptive" instead. I've also got a small 40mm fan running at 7V which blows 1.5-2.0 CFM of air across the heat sink. This dropped the temp considerably.

If you have the right version of OpenSSL, OpenVPN will use AES-GCM mode.

The box should be good for something approaching 8Gbps throughput, when all is said and done. That said, the architecture of OpenVPN might not allow that to happen.

… but IPsec should.

In general, I am also unhappy with how the numbers are generated. So unhappy that I'm funding an outside effort, in the guise of 10Gbps throughput testing for pfSense.

But what we have is … what we have.

stephenw10

Sounds good, I look forward to some updated figures.
I think everyone would agree that the current hardware sizing guidelines are looking pretty stale. I have suggested some manner of community generated database in the past but it never got much traction.

Steve

LOL, I just submitted an update to these yesterday.

"pretty stale" is a kind way of putting it.

"Completely, 100% out of date, and showing an obvious lack of attention." is more accurate.

aus_guy

The form factor of these looks ideal to me but 8 core seems overkill and would mean power consumption would be a bit high aswell.
Im wondering if anyone has seen any major difference between the A1SRi-2758F board used in these boxes to the A1SRi-2558F which is 4-core
From a comparison of the spec sheets they seem identical apart from the core count on cpu but i could be missing something.

I know that this isn't something that's available on the pfsense store and yes i would prefer to buy from pfsense/EFS but nothing in this range is available.

(I'm always amused and amazed when people think we're stupid.)

Power consumption of the C2758, as used on the A1SRI-2758F, is 20W TDP.
Power consumption of the C2558, as used on the A1SRi-2558F, is 15W TDP.

http://ark.intel.com/products/77988/Intel-Atom-Processor-C2758-4M-Cache-2_40-GHz?q=c2758
http://ark.intel.com/products/77983/Intel-Atom-Processor-C2558-2M-Cache-2_40-GHz

Just as a quick-n-dirty comparison, the A1SRi-2558F motherboard is $360 on Amazon

http://www.amazon.com/Supermicro-Mini-DDR3-Motherboards-MBD-A1SRi-2558F-O/dp/B00HS4NLHA

The A1SRI-2758F is $340 on Amazon. Yes, it's less.

http://www.amazon.com/Supermicro-MiniITX-Retail-Motherboards-MBD-A1SRI-2758F-O/dp/B00FM4M7TQ/ref=sr_1_1?s=electronics&ie=UTF8&qid=1400397161&sr=1-1&keywords=A1SRi-2758F

Since everything else in the box will be the same, let's you've spent $40 to save 5W.

There are 8765 hours in a year, yielding a delta slightly less than 24.09 kWh used in a year.

At the highest tier, Austin energy charges residential customers $0.11400 / kWh. At this rate, you
will save $2.74 per year, by spending at least $40 more to get half as many cores.

Perhaps you don't live in Austin, but I did pick the highest residential rate (in Austin).
At that rate, (assumptions noted), assuming electricity rates do not change (more assumptions), and not
counting interest or TVM, it will take you around 14.6 years to pay back.

Of course, you live in Australia, not Austin. Using this report from the Australian Energy Market Commission, the average market offer price for residential ("Household") consumption is between 24.49 and 24.67 c/kWh, but these are Australian 'cents', not US 'cents':
http://www.aemc.gov.au/media/docs/2013-Residential-Electricity-Price-Trends-Final-Report-723596d1-fe66-43da-aeb6-1ee16770391e-0.PDF

This changes the equation only slightly. Making the (entirely unwarranted) assumption that you could purchase the Supermicro board(s) at US prices (and ignoring shipping / duties / etc.), I note that at today's exchange rate, 1.00 USD will buy 1.07 AUD. Picking the 24.49 c/kWh, this equates to 0.23 USD / kWh. This is within striking distance of "2X' the electricity rate, so you might well pay off the difference in price in half the time, or 7.3 years.

While there is a 4-core board coming (shhh!, it's a secret), it is not the A1SRi-2558F.

Edit: noted that aus_guy is in Australia, not Austin. Adjusted / added figures to reflect his situation, sorta.

@asterix:

I haven't tested the C2758s yet but my typical recommendation has always been to go for an i3. There is not much of power saving difference or in other words you won't be a $100 richer after running an atom or other CPU slower than the i3.

Assuming you pick a slow (1.7GHz), dual core, 'mobile' Haswell i3 such as:
http://ark.intel.com/products/75107/Intel-Core-i3-4010U-Processor-3M-Cache-1_70-GHz?q=i3-4010U

The TDP of the part is 15W, .vs 20W for the C2758. You have 2 cores at 1.7GHz, instead of 8 at 2.4GHz.

Assuming you pick a faster i3, such as the i3-4000M, TDP is 37 W, and you only have 2 cores, instead of 8.
http://ark.intel.com/products/75104/Intel-Core-i3-4000M-Processor-3M-Cache-2_40-GHz

Many people do not understand that though the C2000 series is named "Atom", it is, in fact, very unlike the Atom CPUs you know. This one, literally, "goes to 11". In simple tests (such as compiling pfSense), it benchmarks a lot like an 8 core Xeon system from 2009, while consuming far less power.

@stephenw10:

Sounds good, I look forward to some updated figures.
I think everyone would agree that the current hardware sizing guidelines are looking pretty stale. I have suggested some manner of community generated database in the past but it never got much traction.

Steve

We're (finally) working on a documented testing method. This would allow community-generated results to be submitted that could be compared.

stephenw10

Great! :)
I look forward to submitting some numbers from my ancient hardware. ;)

Steve

NTMMFTS

Anyone know if there are any install/driver issues with 2.1.5 or 2.2 with the supermicro c2758 board? I want to build one and just want to be sure of what work I may have to do to get it running. Please share any info you're aware of.

thanks!
Jay.

kroberts

I just ordered a different 2758 board. The intel c2000 processors are oriented toward communications, and the c2*58 processors have QuickAssist. This is hardware acceleration for encryption and compression. These chips are designed to replace larger power hungry servers with smaller atoms in the router/vpn market.

I'm a pfSense n00b but have almost 20 years with Linux. The Linux kernel just got support for quick-assist. Apps don't seem to have it yet. I expect FreeBSD will be at approximately the same state.

The board I ordered supports 64g ecc or non-ecc memory. Surely there's a reason why it supports twice as much as a mainstream i7. Obviously for mainstream uses the atom will be stomped by the i7, but the pfSense use case is exactly what c2*58 processors were designed to do.

@Asterix:

I haven't tested the C2758s yet but my typical recommendation has always been to go for an i3. There is not much of power saving difference or in other words you won't be a $100 richer after running an atom or other CPU slower than the i3. The processor itself slows down when not required and has the added juice to push snort, dansguardian with clamd, squid and other packages all at the same time.

Wot?

i3-4150 is 54W http://ark.intel.com/products/77486/Intel-Core-i3-4150-Processor-3M-Cache-3_50-GHz

C2758 is 20W http://ark.intel.com/products/77988/Intel-Atom-Processor-C2758-4M-Cache-2_40-GHz?q=c2758

Today, there is probably a slight performance advantage to the i3. Tomorrow you're going to want the cores.

Trust me. You are going to want more cores.

@kroberts:

I just ordered a different 2758 board. The intel c2000 processors are oriented toward communications, and the c2*58 processors have QuickAssist. This is hardware acceleration for encryption and compression. These chips are designed to replace larger power hungry servers with smaller atoms in the router/vpn market.

C2*58 doesn't mean QuickAssist. See: http://ark.intel.com/compare/77988,77986,77984,77983,77981,77979,81270,77978,77976,81328

@kroberts:

I'm a pfSense n00b but have almost 20 years with Linux. The Linux kernel just got support for quick-assist. Apps don't seem to have it yet. I expect FreeBSD will be at approximately the same state.

The linux kernel got QAT support last Summer. You will never guess who Intel has tapped to do the QAT driver for FreeBSD.

@kroberts:

The board I ordered supports 64g ecc or non-ecc memory. Surely there's a reason why it supports twice as much as a mainstream i7. Obviously for mainstream uses the atom will be stomped by the i7, but the pfSense use case is exactly what c2*58 processors were designed to do.

Have you priced 64GB of ECC ram? :P