Dual CPU Quad Core Xeon Server for $200
-
Specs: http://unixsurplus.com/product/2u-rack-server-intel-s5000psl-2x-xeon-e5345-quad-core
Picked up this baby on craigslist for $200. although, it only came with (1) 250GB HDD. I ended up paying an extra $10 for a riser card, and $30 for a dual-port PRO1000 nic.
Is it overkill? Absolutely…but for the price, it was less expensive than building a rackable unit with consumer hardware.
Using Comcast Metro-E 30x30, but getting closer to 40Mbps both ways with about ~2ms latency.Requirements:
-40 Users (with voip phones)
-3 Networks, (1 windows domain, 1 mac domain, 1 voip network)
-basic firewall config with traffic shaping, port forwards, blah blah
-bandwithd
-soon to add squid/guard, captive portal
-
Nice box! Looks like it would suck down a lot of power though! I like my little Ivy Bridge i3 box :D
-
If I was paying the electric bill I'd probably pull a physical CPU.
-
Yes, yes I realize it isn't the most practical or power efficient. It pulls about 100w. It's the equivalent of an extra desktop running. I may pull one CPU and see what power saving advantages come of it. Using local utility rates, it comes out to $0.19/day.
-
Yes, yes I realize it isn't the most practical or power efficient. It pulls about 100w. It's the equivalent of an extra desktop running. I may pull one CPU and see what power saving advantages come of it. Using local utility rates, it comes out to $0.19/day.
Oh, don't get me wrong, it looks like a good box and a seems to be a good deal. I just ran a couple old Dell PowerEdge 2850's for a short time 'till I started getting the bills. The Core2Duo+ era Xeons are better on power than my older P4 based models, but I still would have expected a bit more than 100W. I think mine were around 300W each, but they also had 4 or 5 10K RPM SCSI drives, and that doesn't help.
The other benefit you may get from pulling a physical CPU when the CPUs themselves have multiple cores, especially with pfSense, is reducing cache swaps. pfSense is largely non-multi-threaded, a couple packages may run another thread, but the bulk of its work is in single threads. (Now, I'm no mluti-thread guru, but this is how I always understood this) When those processes are scheduled, they can be scheduled across logical CPUs, be them physical sockets, cores, or Hyper-Threading logical CPUs. When a process migrates from one CPU to another, it may have to transfer its cache, often this is very fast, especially when it's on the same socketed CPU, via snooping. But, when it does it across sockets, it has to transfer that over the CPU bus between processors, which is slower.
Of course, it's quite possible that you'd only be able to tell through specific timed tests, or maybe not at all, but there's no reason to waste power on something that's potentially slowing you down, anyway.
-
Dont forget Core2Quads are really 2 Core2Duo's so going from say core 0 to core 2 would require going over the FSB even if they are in the same socket..
-
"$30 for a dual-port PRO1000 nic."
Where did you get a dualport gig nic for $30?? I would love to change out extra nic I put in the my n40l box with dual. And $30 would make it a no brainer to do. Only issue is low profile.
-
Dont forget Core2Quads are really 2 Core2Duo's so going from say core 0 to core 2 would require going over the FSB even if they are in the same socket..
Oh, yeah, I see what you mean, that's bad too. I had been under the impression that there was a cache transport on the die, and, after further review, I'm wrong.
So, it's quite possible that a Core2Duo may be better than a Core2Quad at the same clock speed, or even better than a pair of Quad core Xeon's or pair of Dual Core Xeon's (previous to the Nehalem arcitechture.) All else being equal, of course. Also assuming you'd actually peg any of those CPUs with normal routing or even VPN at 40Mb/s
-
"$30 for a dual-port PRO1000 nic."
Where did you get a dualport gig nic for $30?? I would love to change out extra nic I put in the my n40l box with dual. And $30 would make it a no brainer to do. Only issue is low profile.
You can find 'em on ebay from time to time, but the low profile requirement will make it more difficult for you since the less expensive ones tend to be the older tall cards that had no option for a LP bracket. If a Broadcom chipset works for you sometimes you can find them for cheaper, not sure about the LP brackets, though (the PCB is often small enough.)
-
Dont forget Core2Quads are really 2 Core2Duo's so going from say core 0 to core 2 would require going over the FSB even if they are in the same socket..
Oh, yeah, I see what you mean, that's bad too. I had been under the impression that there was a cache transport on the die, and, after further review, I'm wrong.
So, it's quite possible that a Core2Duo may be better than a Core2Quad at the same clock speed, or even better than a pair of Quad core Xeon's or pair of Dual Core Xeon's (previous to the Nehalem arcitechture.) All else being equal, of course. Also assuming you'd actually peg any of those CPUs with normal routing or even VPN at 40Mb/s
Yeah, this is true. However there is one thing you DONT need to worry about and that is NUMA, as the Core2's still use the chipset-based memory controller. You'd only run into NUMA if you have a multi socket Nehalhem or newer. I think the best Core2 for pfSense would be a high clocked 45nm(Penryn) with the full 6MB of L2 cache, and use PowerD so speed step kicked in.
-
(Now, I'm no mluti-thread guru, but this is how I always understood this)
On the contrary, I see you ARE quite the guru on this subject! :)
http://forum.pfsense.org/index.php/topic,54475.msg294550.html#msg294550
Anyway, next time I have availability to take down the network I'll give it a shot. Do you have any references/links for "tests" that I can run to validate your hypothesis? Also, is there any way to tell whether or not processes are in fact spanning across CPUs?
Your knowledge of hardware is indispensable to many on this forum (myself included).
Thanks! -
"$30 for a dual-port PRO1000 nic."
Where did you get a dualport gig nic for $30?? I would love to change out extra nic I put in the my n40l box with dual. And $30 would make it a no brainer to do. Only issue is low profile.
As others have pointed out, it was an ebay deal. However, after googling the prices of the low profile offerings…seems like it adds considerable costs.
-
(Now, I'm no mluti-thread guru, but this is how I always understood this)
On the contrary, I see you ARE quite the guru on this subject! :)
http://forum.pfsense.org/index.php/topic,54475.msg294550.html#msg294550
Anyway, next time I have availability to take down the network I'll give it a shot. Do you have any references/links for "tests" that I can run to validate your hypothesis? Also, is there any way to tell whether or not processes are in fact spanning across CPUs?
Your knowledge of hardware is indispensable to many on this forum (myself included).
Thanks!I was recalling, but couldn't actually remember where I made that post, now I can update it, thanks. (Not that I made the effort to actually search for it, but now I have no excuse for not updating it.)
Really, though, I'm mostly just good at regurgitating research. I have no need to be right, but I do have a need for the right information to be available. Half of my posts end up being "Hmm, I think I might know, but it'd be good to check it out and find out, and if I've gone so far as to research it, I'll learn it better if I formulate it in to a post."
What I was getting at was, if you're up for testing that, dude, please do! I would be very interested in the results. Feel free to show my hypothesis wrong. We had someone testing real world results of the various virtual nics available in VMWare, was great to have tested info rather than just relying on conjecture (turned out to not make as much difference as expected, but some.)
I don't know BSD well enough to help you with determining if a process is spanning. I saw a post on another forum about a similar subject, a development forum, but the subject seemed to drop before any answers came out.