Discussion:
[pfSense] NIC support
Ryan Clough
2014-10-13 17:17:18 UTC
Permalink
I am looking at the HP ProLiant DL320e Gen8 v2 and having trouble
determining whether or not the hardware is supported by pfSense
2.1.5-RELEASE. I found this thread(
https://forum.pfsense.org/index.php?topic=71523.0) and it seems like I am
going to have trouble with drivers. Here are the three hardware components
that concern me:

HP B120i on-board SATA controller
HP 332i on-board 2-port NIC (Broadcom BCM5720 chipset)
HP 361T PCIe 2-port NIC (Intel I350 chipset)

Can anyone confirm or deny this hardware's compatibility? Anyone out there
running on a Gen8 v2? I do not have the option to run pfSense in a virtual
environment.

Thanks,
Ryan
--
This email and its contents are confidential. If you are not the intended
recipient, please do not disclose or use the information within this email
or its attachments. If you have received this email in error, please report
the error to the sender by return email and delete this communication from
your records.
Vick Khera
2014-10-13 17:39:37 UTC
Permalink
Post by Ryan Clough
I am looking at the HP ProLiant DL320e Gen8 v2 and having trouble
determining whether or not the hardware is supported by pfSense
2.1.5-RELEASE. I found this
thread(https://forum.pfsense.org/index.php?topic=71523.0
) and it seems like I am going to have trouble with drivers. Here are the
In general HP servers work really well with FreeBSD.

When you say "looking" are you in possession of one and need to make
it work, or are you about to buy one? Is there some specific
requirement about that hardware that makes you want to get it over
anything else?

I personally have found that the C2758 sold by both netgate and
pfsense directly to be a spectacularly capable device and it is fairly
priced and includes support. I would recommend that based on what
you've described above unless there's some other special need you
have.
Ryan Clough
2014-10-13 17:50:23 UTC
Permalink
Thanks for your quick reply, Vick. By "looking", I mean I have not yet
purchased one but have a quote from our supplier. I have looked at the
C2758 but one of my requirements is RAID1. Also, I am not comfortable
deploying an SSD that will be running Squid. This router needs to have the
capability to run for years with minimal maintenance. We are planning to
deploy this as part of one of our products. Thanks again for your help.

Ryan Clough
Information Systems
Decision Sciences International Corporation
<http://www.decisionsciencescorp.com/>
<http://www.decisionsciencescorp.com/>
Post by Vick Khera
Post by Ryan Clough
I am looking at the HP ProLiant DL320e Gen8 v2 and having trouble
determining whether or not the hardware is supported by pfSense
2.1.5-RELEASE. I found this
thread(https://forum.pfsense.org/index.php?topic=71523.0
) and it seems like I am going to have trouble with drivers. Here are the
In general HP servers work really well with FreeBSD.
When you say "looking" are you in possession of one and need to make
it work, or are you about to buy one? Is there some specific
requirement about that hardware that makes you want to get it over
anything else?
I personally have found that the C2758 sold by both netgate and
pfsense directly to be a spectacularly capable device and it is fairly
priced and includes support. I would recommend that based on what
you've described above unless there's some other special need you
have.
_______________________________________________
List mailing list
https://lists.pfsense.org/mailman/listinfo/list
--
This email and its contents are confidential. If you are not the intended
recipient, please do not disclose or use the information within this email
or its attachments. If you have received this email in error, please report
the error to the sender by return email and delete this communication from
your records.
Vick Khera
2014-10-13 20:26:32 UTC
Permalink
Post by Ryan Clough
Thanks for your quick reply, Vick. By "looking", I mean I have not yet
purchased one but have a quote from our supplier. I have looked at the C2758
but one of my requirements is RAID1. Also, I am not comfortable deploying an
SSD that will be running Squid. This router needs to have the capability to
run for years with minimal maintenance. We are planning to deploy this as
part of one of our products. Thanks again for your help.
Based on this page
http://blog.hostileadmin.com/2012/06/14/freebsd-on-hp-proliant-dl360p-g8-servers/
I'd say look for something else, unless the two Intel NICs are
sufficient. Further investigation
<https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=182387#c2> shows
BCM5720 is not supported until 8.4 and 9.2. Given pfSense is 8.3 you
will have two NICs that are unusable.
Ulrik Lunddahl
2014-10-14 07:48:54 UTC
Permalink
Post by Vick Khera
In general HP servers work really well with FreeBSD.
When you say "looking" are you in possession of one and need to make it work, or are you about to buy one? Is there some specific requirement about that hardware that makes you want to get it over anything else?
I personally have found that the C2758 sold by both netgate and pfsense directly to be a spectacularly capable device and it is fairly priced and includes support. I would recommend that based on what you've described above unless there's some other special need you have.
I know that:

"- Blistering fast Intel® AtomT Rangeley C2758 8 core SoC This is not your father's Atom!"

Probably is a beast compared to what we normally expect form the Atom range, but to compare it with an up to date Dual Xeon Platform is just not going to make a lot of sense.

Hardware quality on the two boxes is also almost incomparable, both are general-purpose platforms, but from different ends of the scale.

Will A SMB without L3 capable switches, that needs routing between 3-4 local subnets (LAN, SERVERS, WIRELESS/GUEST, OTHER/DMZ) as close to wirespeed as possible, be happy with a C2758. ?


Med venlig hilsen, Best regards
Ulrik Lunddahl

Sales Manager - Salgschef
PROconsult Data A/S - Landbrugsvej 2 - 5260  Odense S
Tel: +45 63113333 - Tel dir: +45 63113341 - Mobil: +45 26363341
E-mail: ***@proconsult.dk - Web site: www.proconsult.dk

    

VSP - Infrastructure Optimization Solutions + VSP - Business Continuity
VTSP - VMware Infrastructure Virtualization + vExpert - 2009, 2010, 2012
VMSP - Veeam Sales Professional + VMTSP - Veeam Technical Sales Professional
Jim Thompson
2014-10-14 12:24:57 UTC
Permalink
Post by Ulrik Lunddahl
Will A SMB without L3 capable switches, that needs routing between 3-4 local subnets (LAN, SERVERS, WIRELESS/GUEST, OTHER/DMZ) as close to wirespeed as possible, be happy with a C2758. ?
Very.

Is a dual socket Xeon a bit faster? Yes.
Does your application need that speed? Unlikely.

Really depends on what you mean by "wirespeed".

-- Jim
Post by Ulrik Lunddahl
Post by Vick Khera
In general HP servers work really well with FreeBSD.
When you say "looking" are you in possession of one and need to make it work, or are you about to buy one? Is there some specific requirement about that hardware that makes you want to get it over anything else?
I personally have found that the C2758 sold by both netgate and pfsense directly to be a spectacularly capable device and it is fairly priced and includes support. I would recommend that based on what you've described above unless there's some other special need you have.
"- Blistering fast Intel® AtomT Rangeley C2758 8 core SoC This is not your father's Atom!"
Probably is a beast compared to what we normally expect form the Atom range, but to compare it with an up to date Dual Xeon Platform is just not going to make a lot of sense.
Hardware quality on the two boxes is also almost incomparable, both are general-purpose platforms, but from different ends of the scale.
Will A SMB without L3 capable switches, that needs routing between 3-4 local subnets (LAN, SERVERS, WIRELESS/GUEST, OTHER/DMZ) as close to wirespeed as possible, be happy with a C2758. ?
Med venlig hilsen, Best regards
Ulrik Lunddahl
Sales Manager - Salgschef
PROconsult Data A/S - Landbrugsvej 2 - 5260 Odense S
Tel: +45 63113333 - Tel dir: +45 63113341 - Mobil: +45 26363341
VSP - Infrastructure Optimization Solutions + VSP - Business Continuity
VTSP - VMware Infrastructure Virtualization + vExpert - 2009, 2010, 2012
VMSP - Veeam Sales Professional + VMTSP - Veeam Technical Sales Professional
_______________________________________________
List mailing list
https://lists.pfsense.org/mailman/listinfo/list
compdoc
2014-10-14 22:15:26 UTC
Permalink
Post by Ulrik Lunddahl
as close to wirespeed as possible, be happy with a C2758. ?
Very
That C2758 has nice specs and should be able to keep up, however there seems to be a throughput problem on at least one brand of board running the C2758. (I think it’s more a problem with the nics than the cpu)



I recently tested various nics and cpus to see if the systems I was building could reach Gigabit Ethernet's max throughput of 1.488Mpps on one port.



Tests were run on AMD FM1+ and AM1 APUs, an FX-4100, and an Intel i5-2400 Sandy Bridge. Tests used the BSD Router Project (BSDRP) OS, and a program named 'pkt-gen'.



During routing tests, I found that an AMD A8-7600 Kaveri was the only cpu I had that was equal in performance to the Intel i5-2400. (the routing tests involved a 3rd test machine, and aren't covered in the scores below)



Anyway, I hope you find this helpful...





In these tests, I used the two fastest test machines connected to each other. One sends, and one receives:



Realtek 8169sc 32-bit PCI card

266935 pps (283752 pkts in 1063001 usec)

Speed: 267.19 Kpps Bandwidth: 128.25 Mbps (raw 179.55 Mbps)



Realtek RTL8111DL, Onboard

405708 pps (406113 pkts in 1000998 usec)

Speed: 404.78 Kpps Bandwidth: 194.29 Mbps (raw 272.01 Mbps)



Intel pro 1000 32-bit PCI card

307102 pps (307586 pkts in 1001577 usec)

Speed: 276.49 Kpps Bandwidth: 132.72 Mbps (raw 185.80 Mbps)



Intel Pro 1000, x1 PCI-e card (no heatsink)

1367299 pps (1453440 pkts in 1063001 usec)

Speed: 1.36 Mpps Bandwidth: 654.85 Mbps (raw 916.79 Mbps)



Intel Pro 1000, x1 PCI-e card, server version (with heatsink)

1488012 pps (1490981 pkts in 1001995 usec)

Speed: 1.49 Mpps Bandwidth: 714.23 Mbps (raw 999.92 Mbps)



Intel PRO/1000 PT, Dual Port, 4x PCI-e, Server Adapter (with heatsink)

1488012 pps (1490981 pkts in 1001995 usec)

Speed: 1.49 Mpps Bandwidth: 714.23 Mbps (raw 999.92 Mbps)





***************************************************************



These tests were using the lowest TDP(watt) APUs I had.

The Intel server nics were the fastest nics tested, and used the least cpu time, so I used those in these tests:



AMD 5150 quad core APU @ 1.6GHz

Intel PRO/1000 PT, Dual Port, 4x PCI-e, Server Adapter (with heatsink)

1179367 pps (1180530 pkts in 1000986 usec)

Speed: 1.17 Mpps Bandwidth: 562.85 Mbps (raw 787.99 Mbps)



AMD 5350 quad core APU @ 2GHz

Intel PRO/1000 PT, Dual Port, 4x PCI-e, Server Adapter (with heatsink)

1488106 pps (1489615 pkts in 1001014 usec)

Speed: 1.48 Mpps Bandwidth: 709.33 Mbps (raw 993.07 Mbps)



AMD 5350 quad APU @ 2GHz

Onboard RTL8111/8168B PCI Express Gigabit Ethernet controller

560938 pps (561565 pkts in 1001117 usec)

Speed: 558.35 Kpps Bandwidth: 268.01 Mbps (raw 375.21 Mbps)



AMD A4-6300 dual core APU @ 3.7GHz

Intel PRO/1000 PT, Dual Port, 4x PCI-e, Server Adapter (with heatsink)

1129784 pps (1130961 pkts in 1001042 usec)

Speed: 1.09 Mpps Bandwidth: 521.00 Mbps (raw 729.39 Mbps)
Jim Thompson
2014-10-15 12:15:16 UTC
Permalink
Post by compdoc
Post by Ulrik Lunddahl
as close to wirespeed as possible, be happy with a C2758. ?
Very
That C2758 has nice specs and should be able to keep up, however there seems to be a throughput problem on at least one brand of board running the C2758.
(I think it’s more a problem with the nics than the cpu)
You seem confused.
Post by compdoc
I recently tested various nics and cpus to see if the systems I was building could reach Gigabit Ethernet's max throughput of 1.488Mpps on one port
Please show your work. Which pkt-gen switches are in use?
Post by compdoc
Tests were run on AMD FM1+ and AM1 APUs, an FX-4100, and an Intel i5-2400 Sandy Bridge.
None of these is the system in question. They don't even run the same cpu.
Post by compdoc
Tests used the BSD Router Project (BSDRP) OS, and a program named 'pkt-gen'.
- I am quite familiar with pkt-gen.

- this list is about pfsense, not the BSDRP
Post by compdoc
During routing tests, I found that an AMD A8-7600 Kaveri was the only cpu I had that was equal in performance to the Intel i5-2400. (the routing tests involved a 3rd test machine, and aren't covered in the scores below)
Pkt-gen does not test routing. What tests did you run?
Post by compdoc
Anyway, I hope you find this helpful...
I don't see where a C2758 is tested.
Post by compdoc
Realtek 8169sc 32-bit PCI card
266935 pps (283752 pkts in 1063001 usec)
Speed: 267.19 Kpps Bandwidth: 128.25 Mbps (raw 179.55 Mbps)
Realtek RTL8111DL, Onboard
405708 pps (406113 pkts in 1000998 usec)
Speed: 404.78 Kpps Bandwidth: 194.29 Mbps (raw 272.01 Mbps)
Intel pro 1000 32-bit PCI card
307102 pps (307586 pkts in 1001577 usec)
Speed: 276.49 Kpps Bandwidth: 132.72 Mbps (raw 185.80 Mbps)
Intel Pro 1000, x1 PCI-e card (no heatsink)
1367299 pps (1453440 pkts in 1063001 usec)
Speed: 1.36 Mpps Bandwidth: 654.85 Mbps (raw 916.79 Mbps)
Intel Pro 1000, x1 PCI-e card, server version (with heatsink)
1488012 pps (1490981 pkts in 1001995 usec)
Speed: 1.49 Mpps Bandwidth: 714.23 Mbps (raw 999.92 Mbps)
Intel PRO/1000 PT, Dual Port, 4x PCI-e, Server Adapter (with heatsink)
1488012 pps (1490981 pkts in 1001995 usec)
Speed: 1.49 Mpps Bandwidth: 714.23 Mbps (raw 999.92 Mbps)
***************************************************************
These tests were using the lowest TDP(watt) APUs I had.
APUs? I thought we were talking C2758.
Post by compdoc
Intel PRO/1000 PT, Dual Port, 4x PCI-e, Server Adapter (with heatsink)
1179367 pps (1180530 pkts in 1000986 usec)
Speed: 1.17 Mpps Bandwidth: 562.85 Mbps (raw 787.99 Mbps)
AMD CPU. NON-identified NIC.
Post by compdoc
Intel PRO/1000 PT, Dual Port, 4x PCI-e, Server Adapter (with heatsink)
1488106 pps (1489615 pkts in 1001014 usec)
Speed: 1.48 Mpps Bandwidth: 709.33 Mbps (raw 993.07 Mbps)
AMD CPU. NON-identified NIC.
Post by compdoc
Onboard RTL8111/8168B PCI Express Gigabit Ethernet controller
560938 pps (561565 pkts in 1001117 usec)
Speed: 558.35 Kpps Bandwidth: 268.01 Mbps (raw 375.21 Mbps)
AMD CPU. NON-identified NIC.
Post by compdoc
Intel PRO/1000 PT, Dual Port, 4x PCI-e, Server Adapter (with heatsink)
1129784 pps (1130961 pkts in 1001042 usec)
Speed: 1.09 Mpps Bandwidth: 521.00 Mbps (raw 729.39 Mbps
AMD CPU. NON-identified NIC.

Now the track has been completely lost.

Jim
compdoc
2014-10-15 15:06:39 UTC
Permalink
When I speak of the C2758, I speak of the product sold at the pfSense store,
as sold by the pfSense store, not the generic pfsense release running on "some
I was speaking of a C2758 board that was tested by someone else, and which wasn’t able to reach Ethernet's maximum throughput. Clearly not all C2758 boards are the same. Buyer beware.



If you have tests results that prove the product you mentioned doesn’t have this problem, feel free to post them. I'd love to see.
You seem confused.
Not at all. You seem defensive.
- this list is about pfsense, not the BSDRP
Never said it was. BSDRP is a tool to test hardware. If the hardware cannot achieve maximum throughput, then pfSense cannot achieve maximum throughput.
Pkt-gen does not test routing. What tests did you run?
Here's a clue: BSD *Router* Project. I doubt you’ve done this sort of testing, so I'm not going to spoil this learning opportunity for you...



However, I will mention one thing: if you try to route 1.488M packets per second through the 'generic' pfSense, it will crash after a minute or so. (and that's not a criticism of pfSense)
I don't see where a C2758 is tested.
I clearly stated what I was testing and how. You seem confused. The OP was asking what hardware might serve his purpose. I offered suggestions.



You're welcome to prove anything I've said was wrong - but with actual test results, and without the misplaced rancor.



Also, it's better to reply to the list, and not send emails directly to me.
Jim Thompson
2014-10-15 19:39:02 UTC
Permalink
-- Jim
Post by compdoc
When I speak of the C2758, I speak of the product sold at the pfSense store,
as sold by the pfSense store, not the generic pfsense release running on "some
I was speaking of a C2758 board that was tested by someone else, and which wasn’t able to reach Ethernet's maximum throughput. Clearly not all C2758 boards are the same. Buyer beware.
If you have tests results that prove the product you mentioned doesn’t have this problem, feel free to post them. I'd love to see.
You seem confused.
Not at all. You seem defensive.
- this list is about pfsense, not the BSDRP
Never said it was. BSDRP is a tool to test hardware.
Actually it's not. Olivier uses it in his work at Orange.
There has been some testing using BSDRP, but it is not "a tool to test hardware".
Post by compdoc
If the hardware cannot achieve maximum throughput, then pfSense cannot achieve maximum throughput.
This is a true statement but it ignores the reality that software also plays a part.
Post by compdoc
Pkt-gen does not test routing. What tests did you run?
Here's a clue: BSD *Router* Project. I doubt you’ve done this sort of testing, so I'm not going to spoil this learning opportunity for you...
You seem defensive.

You were testing forwarding, by the look of it. This is not all there is to routing. I will not further ecludiate because you are obviously an expert.

While you "doubt" we "have done this sort of testing" you should look at: https://github.com/gvnn3/conductor

Quoting README

[...]
A common use for Conductor is to test a network devices, such as a router or firewall, that is connected to multiple senders and receivers. Each of the senders, receivers, and the device under test
(DUT) are a Player, and another system is designated as the Conductor.

[...]

This work supported by: Rubicon Communications, LLC (Netgate)
Conductor uses pkt-gen or iperf, though our preference going forward is pit-gen. Recent additions to pkt-gen include playback of pcap files, for more repeatable testing. It's also important to be able to test multiple senders and receivers. I will not further ecludiate because you are an expert.
Post by compdoc
However, I will mention one thing: if you try to route 1.488M packets per second through the 'generic' pfSense, it will crash after a minute or so. (and that's not a criticism of pfSense)
That's an interesting result. We've not seen it.
Which particular hardware were you using?
Which version of pfsense?
Any tunables?
What switches to pkt-gen?
Post by compdoc
I don't see where a C2758 is tested.
I clearly stated what I was testing and how. You seem confused. The OP was asking what hardware might serve his purpose. I offered suggestions.
You're welcome to prove anything I've said was wrong - but with actual test results, and without the misplaced rancor.
Also, it's better to reply to the list, and not send emails directly to me.
_______________________________________________
List mailing list
https://lists.pfsense.org/mailman/listinfo/list
compdoc
2014-10-15 21:06:19 UTC
Permalink
Post by Jim Thompson
There has been some testing using BSDRP, but it is not "a tool to test hardware".
I used it as a tool to benchmark my hardware. There are several examples on their website of using it for just that purpose.



It also a tool to build simple routers. It has very little overhead and runs on freebsd, which made it interesting.



It is a tool.
Post by Jim Thompson
You were testing forwarding, by the look of it. This is not all there is to routing.
The testing results I posted were pure packets per second without forwarding. I also tested forwarding but did not post the results, and I mentioned that.
Post by Jim Thompson
Post by compdoc
However, I will mention one thing: if you try to route 1.488M packets per second through the 'generic' pfSense, it will crash after a minute or so.
That's an interesting result. We've not seen it.
These crashes happened during a forwarding test using pfSense. I disabled packet filtering to try to lessen overhead, but it doesn’t seem that pfSense is designed to push a great flood of very tiny packets for any length of time, in one interface and out another.



And I don’t fault it for that. For normal types of traffic, it’s a very capable firewall. It would be interesting to know your results.
Jim Thompson
2014-10-15 21:16:24 UTC
Permalink
Post by compdoc
Post by Jim Thompson
There has been some testing using BSDRP, but it is not "a tool to test hardware".
I used it as a tool to benchmark my hardware. There are several examples on their website of using it for just that purpose.
I am well-aware of Olivier’s work in this area, as are many in the FreeBSD community.
Post by compdoc
Post by Jim Thompson
You were testing forwarding, by the look of it. This is not all there is to routing.
The testing results I posted were pure packets per second without forwarding. I also tested forwarding but did not post the results, and I mentioned that.
So this (“pure packets per second without forwarding”) reduces to just “testing netmap”.
Post by compdoc
Post by Jim Thompson
Post by compdoc
However, I will mention one thing: if you try to route 1.488M packets per second through the 'generic' pfSense, it will crash after a minute or so.
That's an interesting result. We've not seen it.
These crashes happened during a forwarding test using pfSense. I disabled packet filtering to try to lessen overhead, but it doesn’t seem that pfSense is designed to push a great flood of very tiny packets for any length of time, in one interface and out another.
And I don’t fault it for that. For normal types of traffic, it’s a very capable firewall. It would be interesting to know your results.
You’re still assigning fault to pfSense, haven’t properly documented what you’re seeing (thus your assertion that this is pfSense, rather than something in your hardware or in
the testing environment) is not well-supported) and haven’t even answered my questions asking for more detail.

I am also well-aware of the performance issues with pf. We’re working on it. You may have missed the blog post yesterday (https://blog.pfsense.org/?p=1473 <https://blog.pfsense.org/?p=1473>).

Jim
compdoc
2014-10-15 22:01:28 UTC
Permalink
Post by Jim Thompson
I am well-aware of Olivier’s work in this area, as are many in the FreeBSD community.
You’ve failed to disprove anything I've said, even the part about tools.
Post by Jim Thompson
You’re still assigning fault to pfSense
Not at all. But it would be nice if any of this pleasant banter becomes useful by pushing someone to actually try this type of testing, to find out why it happens. And if not, oh well...



By the way, does the C2758 hardware sold by pfSense include pps performance information? Has anyone with this hardware tested it? (speaking to others who might be reading this)



You suggest it can operate at near 'wirespeed', or at least that the OP will be very happy with a C2758 , but you’ve not proven it.
Jim Thompson
2014-10-16 04:55:47 UTC
Permalink
Post by compdoc
Post by Jim Thompson
I am well-aware of Olivier’s work in this area, as are many in the FreeBSD community.
You’ve failed to disprove anything I've said, even the part about tools.
I'm not going to argue with an individual who defines terms to suit his position.
Post by compdoc
Post by Jim Thompson
You’re still assigning fault to pfSense
Not at all.
I see. "It crashes!", but no detail forthcoming.
Post by compdoc
But it would be nice if any of this pleasant banter becomes useful by pushing someone to actually try this type of testing, to find out why it happens. And if not, oh well...
We're not quite ready to publish the results, (because we want people to be able to reproduce them, and maybe put an end to this "benchmarking as a sport"), but yes, the testing is certainly taking place.
Post by compdoc
By the way, does the C2758 hardware sold by pfSense include pps performance information? Has anyone with this hardware tested it? (speaking to others who might be reading this)
You suggest it can operate at near 'wirespeed', or at least that the OP will be very happy with a C2758 , but you’ve not proven it.
There is no proof, except that which is documented and reproducible. We're doing something like science here.

Jim
compdoc
2014-10-16 07:06:13 UTC
Permalink
Post by Jim Thompson
I am well-aware of Olivier’s work in this area, as are many in the FreeBSD community.
There is no proof, except that which is documented and reproducible. We're doing something like science here.
Hmm, proof. Well, maybe a scientist like yourself can appreciate my concern over this direct quote from the BSD Router Project, of which you are so well-aware:



"Intel Rangeley: Atom C2758 (8 cores) at 2.4GHz"

"Embedded Intel i354 4-port gigabit Ethernet"

"8Gb of RAM"

"Debugging slow throughput in progress
"

"With the default value of igb(4) drivers that use all 8 cores, this system is not able to received more than 585Kpps (far from the gigabit line-rate 1.488Mpps) on one port ?!?!"

"Last modified: 2014/03/13 20:16 by olivier"





As I said in my original post, I'm know the C2758 is capable according to its specs, however buyer beware...
Jim Thompson
2014-10-16 13:19:22 UTC
Permalink
Post by compdoc
Post by Jim Thompson
I am well-aware of Olivier’s work in this area, as are many in the FreeBSD community.
There is no proof, except that which is documented and reproducible. We're doing something like science here.
"Intel Rangeley: Atom C2758 (8 cores) at 2.4GHz"
"Embedded Intel i354 4-port gigabit Ethernet"
"8Gb of RAM"
"Debugging slow throughput in progress
"
"With the default value of igb(4) drivers that use all 8 cores, this system is not able to received more than 585Kpps (far from the gigabit line-rate 1.488Mpps) on one port ?!?!"
"Last modified: 2014/03/13 20:16 by olivier"
As I said before, I am aware of Olivier's work. That you are "concerned" is understandable, but also immaterial, as it is clear from this thread that your understanding of the issues, tools(!), terms of art and resolutions is limited.

The concern I have is not your lack of understanding. We all lack knowledge. It's what comes next that marks the difference between progress and the "crabs in a bucket" mentality that often impedes progress.

Here, you perform an act commonly known as "I read it on the Internet" (so it must be true.)

The difference between Olivier's setup and ours (assuming pfsense 2.1.1+), is tuning. It's well-understood that the default install isn't optimal. We addressed this earlier in the year.

Since then we've been concentrating more on a proper test infrastructure, (Conductor), support for AES-GCM mode for IPSec, (with support for AES-NI acceleration), and measuring the performance of "pf" with the on-chip performance counters.

The first result of the pf performance work is an improved (at least 9% faster with 95% confidence) hash function for pf.

A second result (not yet available in pfSense as it requires work from FreeBSD -HEAD) yields another 25% improvement compared to the stock pf in 10.0/10.1.

Work continues.
Post by compdoc
As I said in my original post, I'm know the C2758 is capable according to its specs, however buyer beware...
Again with the insult and denigration. Do you own a C2758?


Jim
compdoc
2014-10-16 16:14:43 UTC
Permalink
Post by Jim Thompson
The difference between Olivier's setup and ours (assuming pfsense 2.1.1+), is tuning
The only way to prove what you say is with numbers. Tuning pfSense won't fix this hardware problem, *if* it exists in your boards.
Post by Jim Thompson
Post by compdoc
As I said in my original post, I'm know the C2758 is capable according to its specs, however buyer beware...
Again with the insult and denigration.
Is it an insult that I think Intel's cpu is capable? Or is it that I suggest a person be cautious when buying these products?
Post by Jim Thompson
That you are "concerned" is understandable, but also immaterial,
as it is clear from this thread that your understanding of the issues,
tools(!), terms of art and resolutions is limited.
...
Here, you perform an act commonly known as "I read it on the Internet" (so it must be true.)
This is a much better example of "insult and denigration". You don’t know me, my methods, or my thinking.
Post by Jim Thompson
Do you own a C2758?
Have you actually bothered to read anything I've said in this conversation?



It's time to end this nonsense. Prove what you say, or shut up.
Andy Holzrichter
2014-10-16 16:42:50 UTC
Permalink
I mostly lurk on this mailing list for the informative discussions, and while this thead is amusing to follow, do you realize who you’re arguing with compdoc? Have you looked at the last part of his email address? If Jim tells us his version of that hardware will do it, I’ll take his word for it barring someone having real proof otherwise. Maybe you need to get one of his boards and run some real tests on it, then report back to the list with what you found.

From: List [mailto:list-***@lists.pfsense.org] On Behalf Of compdoc
Sent: Thursday, October 16, 2014 11:15 AM
To: 'pfSense Support and Discussion Mailing List'
Subject: Re: [pfSense] NIC support
Post by Jim Thompson
The difference between Olivier's setup and ours (assuming pfsense 2.1.1+), is tuning
The only way to prove what you say is with numbers. Tuning pfSense won't fix this hardware problem, *if* it exists in your boards.
Post by Jim Thompson
Post by compdoc
As I said in my original post, I'm know the C2758 is capable according to its specs, however buyer beware...
Again with the insult and denigration.
Is it an insult that I think Intel's cpu is capable? Or is it that I suggest a person be cautious when buying these products?
Post by Jim Thompson
That you are "concerned" is understandable, but also immaterial,
as it is clear from this thread that your understanding of the issues,
tools(!), terms of art and resolutions is limited.
...
Here, you perform an act commonly known as "I read it on the Internet" (so it must be true.)
This is a much better example of "insult and denigration". You don’t know me, my methods, or my thinking.
Post by Jim Thompson
Do you own a C2758?
Have you actually bothered to read anything I've said in this conversation?

It's time to end this nonsense. Prove what you say, or shut up.
compdoc
2014-10-16 17:45:38 UTC
Permalink
Post by Andy Holzrichter
do you realize who you’re arguing with compdoc?
Yeah, I'm arguing with a guy that not only attacked me for suggesting a person be careful about buying certain hardware, he also attacked the work of Olivier from BSDRP.
Jim Thompson
2014-10-17 00:47:27 UTC
Permalink
Post by compdoc
Post by Andy Holzrichter
do you realize who you’re arguing with compdoc?
Yeah, I'm arguing with a guy that not only attacked me for suggesting a person be careful about buying certain hardware, he also attacked the work of Olivier from BSDRP.
I never attacked Olivier. I have a ton of respect both for him and BSDRP.

Jim
compdoc
2014-10-17 22:54:43 UTC
Permalink
I wanted to add one more thing. Maybe this will help avoid future misunderstandings...
"Will A SMB without L3 capable switches, that needs routing between 3-4 local subnets (LAN, SERVERS, WIRELESS/GUEST, OTHER/DMZ) as close to wirespeed as possible, be happy with a C2758. ?"
Now, I realize that the vast majority of users and businesses in the world don’t need a wirespeed router, and they have no idea what one is. Their internet connections just aren't fast enough to require one, and they don’t use them internally.



The fact that Ulrik was asking this question means that he not only knows what one is, but he has a specific requirement.



I've seen others asking this same question on IRC but with a different requirement: they were getting Google Fiber connections and they knew enough to want a server powerful enough to take full advantage of the connection. One guy I saw chose a system with fairly expensive dual Xeon cpus. I thought he was crazy.



Their questions made me curious, and I decided to see just which hardware I had on hand could reach gigabit line-rates. (pkt-gen measures this bandwidth as 714.23 Mbps (raw 999.92 Mbps), at 1.488Mpps)



I was surprised at the results. Nics connected to the PCI bus were dogs. Nics connected to the PCI-e bus were lots faster, and some could reach 1.488Mpps. Also, nics with 4 pci-e lanes were faster than nics with 1 pci-e lane.



Furthermore, I found that to forward packets at 1.488Mpps requires not only a fast NIC, but also a cpu that was capable of pushing traffic through that fast.



The only cpus I had on hand there were capable, was an Intel i5, and a newly released Amd Kaveri APU. (with Steamroller cores)



Anyway, Ulrik asked if he'd be happy with a C2758, and I had read on the BSD-RP site that the C2758 board they were testing wasn’t capable of 1.488Mpps. It was about half that, even though it had Intel based nics.



And while that’s still blazing fast, I felt it might not be fast enough for the knowledgeable people asking these questions.



It would be a shame for anyone to buy something so expensive and expecting certain results, and not getting them.



Even a cheap 5 port gigabit switch can forward traffic at 1.488Mpps, so if the devices sold by pfSense and elsewhere are capable of full wirespeed, then those devices would be an excellent buy.



More so, because of the tuned software and support they'd be getting along with it.



compdoc
Jim Thompson
2014-10-18 06:21:02 UTC
Permalink
So,

The only people getting a google fiber connection *today* live in Provo, UT or Kansas City.

Google Funer is being built out in Austin, but won't be available until early 2015. My neighborhood will get it in the second pass, but I have a Grande 1Gbps/1Gbps connection to my house today, and Grande terminates in the data center next to pfSense World HQ. (We have a 10Gbps fiber connection to our cabinet there.)

So I have a <10ms RTT 1Gbps path from home to work, today. In the next couple months, I'll have two. :-)

Neither pfSense or FreeBSD will forward at 1.488Mpps on a C2758 today, but running the l3fwd app from DPDK on a 2 core C2758 CPU fitted with a dual port 10Gbps card will run at 14.88Mpps.

https://github.com/Pktgen/Pktgen-DPDK/tree/master/dpdk/examples/l3fwd

(And it's trivial to make 1.488 happen in the igb ports. Don't go there.)

A simple bridge over netmap will yield the same result. (With pkt-gen running on either side.)

So the problem is not (as you assert) in the hardware, but rather, in the FreeBSD (and, honestly Linux too) stack(s).

But I've already explained that we're working on it.

-- Jim
Post by compdoc
I wanted to add one more thing. Maybe this will help avoid future misunderstandings...
"Will A SMB without L3 capable switches, that needs routing between 3-4 local subnets (LAN, SERVERS, WIRELESS/GUEST, OTHER/DMZ) as close to wirespeed as possible, be happy with a C2758. ?"
Now, I realize that the vast majority of users and businesses in the world don’t need a wirespeed router, and they have no idea what one is. Their internet connections just aren't fast enough to require one, and they don’t use them internally.
The fact that Ulrik was asking this question means that he not only knows what one is, but he has a specific requirement.
I've seen others asking this same question on IRC but with a different requirement: they were getting Google Fiber connections and they knew enough to want a server powerful enough to take full advantage of the connection. One guy I saw chose a system with fairly expensive dual Xeon cpus. I thought he was crazy.
Their questions made me curious, and I decided to see just which hardware I had on hand could reach gigabit line-rates. (pkt-gen measures this bandwidth as 714.23 Mbps (raw 999.92 Mbps), at 1.488Mpps)
I was surprised at the results. Nics connected to the PCI bus were dogs. Nics connected to the PCI-e bus were lots faster, and some could reach 1.488Mpps. Also, nics with 4 pci-e lanes were faster than nics with 1 pci-e lane.
Furthermore, I found that to forward packets at 1.488Mpps requires not only a fast NIC, but also a cpu that was capable of pushing traffic through that fast.
The only cpus I had on hand there were capable, was an Intel i5, and a newly released Amd Kaveri APU. (with Steamroller cores)
Anyway, Ulrik asked if he'd be happy with a C2758, and I had read on the BSD-RP site that the C2758 board they were testing wasn’t capable of 1.488Mpps. It was about half that, even though it had Intel based nics.
And while that’s still blazing fast, I felt it might not be fast enough for the knowledgeable people asking these questions.
It would be a shame for anyone to buy something so expensive and expecting certain results, and not getting them.
Even a cheap 5 port gigabit switch can forward traffic at 1.488Mpps, so if the devices sold by pfSense and elsewhere are capable of full wirespeed, then those devices would be an excellent buy.
More so, because of the tuned software and support they'd be getting along with it.
compdoc
_______________________________________________
List mailing list
https://lists.pfsense.org/mailman/listinfo/list
Jim Thompson
2014-10-18 06:36:29 UTC
Permalink
Corrections inline.

I blame beer.

-- Jim
Post by Jim Thompson
So,
The only people getting a google fiber connection *today* live in Provo, UT or Kansas City.
Google Funer
Fiber.
Post by Jim Thompson
is being built out in Austin, but won't be available until early 2015. My neighborhood will get it in the second pass, but I have a Grande 1Gbps/1Gbps connection to my house today, and Grande terminates in the data center next to pfSense World HQ. (We have a 10Gbps fiber connection to our cabinet there.)
So I have a <10ms RTT 1Gbps path from home to work, today. In the next couple months, I'll have two. :-)
Neither pfSense or FreeBSD will forward at 1.488Mpps on a C2758 today, but running the l3fwd app from DPDK on a 2
8
Post by Jim Thompson
core C2758 CPU fitted with a dual port 10Gbps card will run at 14.88Mpps.
https://github.com/Pktgen/Pktgen-DPDK/tree/master/dpdk/examples/l3fwd
(And it's trivial to make 1.488 happen in the igb ports. Don't go there.)
A simple bridge over netmap will yield the same result. (With pkt-gen running on either side.)
So the problem is not (as you assert) in the hardware, but rather, in the FreeBSD (and, honestly Linux too) stack(s).
But I've already explained that we're working on it.
-- Jim
Post by compdoc
I wanted to add one more thing. Maybe this will help avoid future misunderstandings...
"Will A SMB without L3 capable switches, that needs routing between 3-4 local subnets (LAN, SERVERS, WIRELESS/GUEST, OTHER/DMZ) as close to wirespeed as possible, be happy with a C2758. ?"
Now, I realize that the vast majority of users and businesses in the world don’t need a wirespeed router, and they have no idea what one is. Their internet connections just aren't fast enough to require one, and they don’t use them internally.
The fact that Ulrik was asking this question means that he not only knows what one is, but he has a specific requirement.
I've seen others asking this same question on IRC but with a different requirement: they were getting Google Fiber connections and they knew enough to want a server powerful enough to take full advantage of the connection. One guy I saw chose a system with fairly expensive dual Xeon cpus. I thought he was crazy.
Their questions made me curious, and I decided to see just which hardware I had on hand could reach gigabit line-rates. (pkt-gen measures this bandwidth as 714.23 Mbps (raw 999.92 Mbps), at 1.488Mpps)
I was surprised at the results. Nics connected to the PCI bus were dogs. Nics connected to the PCI-e bus were lots faster, and some could reach 1.488Mpps. Also, nics with 4 pci-e lanes were faster than nics with 1 pci-e lane.
Furthermore, I found that to forward packets at 1.488Mpps requires not only a fast NIC, but also a cpu that was capable of pushing traffic through that fast.
The only cpus I had on hand there were capable, was an Intel i5, and a newly released Amd Kaveri APU. (with Steamroller cores)
Anyway, Ulrik asked if he'd be happy with a C2758, and I had read on the BSD-RP site that the C2758 board they were testing wasn’t capable of 1.488Mpps. It was about half that, even though it had Intel based nics.
And while that’s still blazing fast, I felt it might not be fast enough for the knowledgeable people asking these questions.
It would be a shame for anyone to buy something so expensive and expecting certain results, and not getting them.
Even a cheap 5 port gigabit switch can forward traffic at 1.488Mpps, so if the devices sold by pfSense and elsewhere are capable of full wirespeed, then those devices would be an excellent buy.
More so, because of the tuned software and support they'd be getting along with it.
compdoc
_______________________________________________
List mailing list
https://lists.pfsense.org/mailman/listinfo/list
_______________________________________________
List mailing list
https://lists.pfsense.org/mailman/listinfo/list
Jim Thompson
2014-10-16 17:23:30 UTC
Permalink
Post by compdoc
Post by Jim Thompson
The difference between Olivier's setup and ours (assuming pfsense 2.1.1+), is tuning
The only way to prove what you say is with numbers. Tuning pfSense won't fix this hardware problem, *if* it exists in your boards.
Your assumption (that there is a hardware problem) is unwarranted. The problem is that FreeBSD (especially FreeBSD 8.3, upon which the current “release” version
of pfSense software (v2.1.5) is based), is not well-tuned to multi-core hardware. We took certain steps to fix the problem (as well as it can be fixed on 8.3) and are working
to improve the situation for both FreeBSD and pfSense. (FreeBSD 10 is better than 8.3, but, as Olivier also discovered, imperfect.)

There is a lot of work to do in this area, including enabling RSS (for forwarding, there is recent work for reception in FreeBSD -HEAD), thread pinning,
additional work on a per-core copy of the state table, more work on flow-table, etc.

It’s all roughly planned, and the subject of some discussion while I have all the pfSense coreteam in Austin this week to discuss this, and what we’re going to do
after the 2.2 release of pfSense.
Post by compdoc
Post by Jim Thompson
Post by compdoc
As I said in my original post, I'm know the C2758 is capable according to its specs, however buyer beware...
Again with the insult and denigration.
Is it an insult that I think Intel's cpu is capable? Or is it that I suggest a person be cautious when buying these products?
Is your position that you are unaware of the meaning of “Caveat emptor”, and it’s history in both English common law and statutory law in all 50 United States?
(Apologies to readers outside the US, but OP is based in Denver, CO, so the point stands.)

You might wish to perform an Internet search for “buyer beware” and see the type of thing that comes up, and then reconsider my reaction in light of same.

You may also wish to review "Laidlaw v. Organ, 15 U.S. 178 (1817)” if you still don’t know what I’m talking about.
Your noisy attempts at persuasion of the consumer base actually require the vendor (that’s me) to respond.
(Never mind the whole “silence is assent” attitude that many hold.)

You gave some results of some tests you performed on an AMD A8-7600 and an i5-2400. I asked for additional details, and you refused to provide any.

You asserted that pfSense crashes under load. (You reported that this “was tested by someone else”) I asked for details, and you refused to provide any.

You asserted that BSDRP is a “tool to test hardware”. You stated that it “has very little overhead and runs on freebsd.”

The reality is that BSDRP is a slightly customized distribution of FreeBSD, it doesn’t “run on FreeBSD”, it *is* FreeBSD, as packaged by Olivier to suit his
purposes at Orange. This is a good thing. That you’ve repurposed it to “test your hardware” is also fine, but your assertion that BSDRP is “a tool to test hardware”
is still false.

Many people use screwdrivers as levers. This doesn’t mean that their usage is correct, nor does it make “a screwdriver is a tool to open paint cans” true.
Post by compdoc
Post by Jim Thompson
Do you own a C2758?
Have you actually bothered to read anything I've said in this conversation?
It's time to end this nonsense. Prove what you say, or shut up.
Fair warning: Being rude will eventually get you removed from the list.

Published numbers are forthcoming, as soon as we’re ready to make the results public. I’ve already exposed the tools we’re using, and some of the improvements we’ve seen.
There is a long history in the project of people making-up benchmark numbers to suit their agenda. There is also a long history in the project of people posting ‘fixes’ for various
issues, including performance issues, where these ‘fixes’ have nothing to do with the actual issue.

The number of times I’ve seen recommendations to "sysctl -w kern.ipc.maxsockbuf=<huge number>” or to set the TCP/UDP default buffer sizes, or set window scaling in an attempt
to increase forwarding performance through ‘pf' makes me cringe. (recent reference: https://forum.pfsense.org/index.php?topic=71949.0)

There are a number of things currently in pfSense that do not lend to absolute performance. mbuf tags and ALTQ are two examples. ALTQ is about a 10% impact on PPS performance.
mbuf tags are the work of the devil. FreeBSD’s penchant for looking up the ARP entry for every single packet (even though it just looked up the ARP entry for the last packet, which was to the same destination) is also a problem. There are some great results from Luigi Rizzo (actual author of the pkt-gen tool) on putting ipfw (the competing packet filter in FreeBSD) over netmap, reaching 7-10Mpps. We will explore pf over netmap (again, after we get pfSense 2.2 released), and hope for similar results.

The point is, we’re focused on it (especially after we get pfSense 2.2 released, such that work we do on pfSense can be taken back “upstream” (to FreeBSD)).

There is also work to do on some of the drivers. Both igb(4) and igbe(4) have issues. cxgbe(4) is in really good shape, by comparison. I don’t want to bother with cxgb(4).
All of the RSS focus in FreeBSD right now is also on igb(4), igbe(4) and cxgbe(4). Even saying this much, someone in the crowd is sure to make the assertion that,
“pfSense isn’t going to support (my favorite driver)!”, which isn’t true, but (favorite driver) might be late to the party. Every statement I make has a certain risk of being taken out of context
to “prove” someone else’s point.

Jim
Ulrik Lunddahl
2014-10-15 07:59:03 UTC
Permalink
Will A SMB without L3 capable switches, that needs routing between 3-4 local subnets (LAN, SERVERS, WIRELESS/GUEST, OTHER/DMZ) as close to wirespeed as possible, be happy with a C2758. ?

Very.

Is a dual socket Xeon a bit faster? Yes.
Does your application need that speed? Unlikely.

Really depends on what you mean by "wirespeed".

The case I always seem to run into is Clients on the LAN, moving a bulk amount of data to/from NAS devices on the SERVER or DMZ subnet, that is typically backup data or data that are somewhat being replicated.

I work a lot with companies dealing in media, and RAW images and/or video is very huge, and devices to store it on is dead cheap.

I also work a lot with virtual environments; backup and replication of virtual machines also generate huge files, which need to be transferred as fast as possible.

So having a hardware router that can both handle internet access from the many LAN clients, and hours of forwarding at interface speed between a few other interfaces is what I would like.

Let’s say that we have a Intel Rangeley Atom 8-core C2758 box with 5 interfaces. (WAN, LAN, SERVERS, OPT1, OPT2)

Will it be able to handle forwarding the packets generated from copying approx. 1 TB of files from LAN to SERVERS and OPT1 to OPT2, and services 50 computers + 50 phones with heavy internet usage.

NAT only, very few rules. ?

I ask because I have no idea how powerful the new Atom’s is.


- Ulrik Lunddahl
Chris L
2014-10-15 08:45:56 UTC
Permalink
Post by Ulrik Lunddahl
Will A SMB without L3 capable switches, that needs routing between 3-4 local subnets (LAN, SERVERS, WIRELESS/GUEST, OTHER/DMZ) as close to wirespeed as possible, be happy with a C2758. ?
Very.
Is a dual socket Xeon a bit faster? Yes.
Does your application need that speed? Unlikely.
Really depends on what you mean by "wirespeed".
The case I always seem to run into is Clients on the LAN, moving a bulk amount of data to/from NAS devices on the SERVER or DMZ subnet, that is typically backup data or data that are somewhat being replicated.
I work a lot with companies dealing in media, and RAW images and/or video is very huge, and devices to store it on is dead cheap.
I also work a lot with virtual environments; backup and replication of virtual machines also generate huge files, which need to be transferred as fast as possible.
So having a hardware router that can both handle internet access from the many LAN clients, and hours of forwarding at interface speed between a few other interfaces is what I would like.
Let’s say that we have a Intel Rangeley Atom 8-core C2758 box with 5 interfaces. (WAN, LAN, SERVERS, OPT1, OPT2)
Will it be able to handle forwarding the packets generated from copying approx. 1 TB of files from LAN to SERVERS and OPT1 to OPT2, and services 50 computers + 50 phones with heavy internet usage.
NAT only, very few rules. ?
I ask because I have no idea how powerful the new Atom’s is.
My first thoughts are:

What is the threat profile you are facing in your organization? Why do you need a firewall between your users and your NAS?

I, personally, would not put pfSense in that duty. If firewalling was not necessary, I’d use a layer 3 switch. And with only 100 devices plus a few servers, I’d wonder why layer 2 wouldn’t suffice.
Loading...