Even a 1,8 ghz Atom can max out a 100mbit line … you don't need much …

niftybunny abuse@to-surf-and-protect.net

Where ignorance is bliss, 'Tis folly to be wise.

Thomas Gray

On 20 Mar 2017, at 21:45, Farid Joubbi joubbi@kth.se wrote:

I'm a little surprised that this kind of question is not documented anywhere, or at least I have not been able to find it. I have only found others asking the same question as me without getting a good answer. There was a quite exhaustive discussion seven years ago: https://lists.torproject.org/pipermail/tor-relays/2010-August/000164.html https://lists.torproject.org/pipermail/tor-relays/2010-August/000164.html

This one got closed before a good answer: https://tor.stackexchange.com/questions/9966/system-requirements-of-tor-rela... https://tor.stackexchange.com/questions/9966/system-requirements-of-tor-relay-how-powerful-a-cpu-and-ram-should-i-use

I am prepared to pay a pile of cash for a computer that I will dedicate to run a relay. I have a 100 Mbit/s symmetric connection. What should I buy? I don't want to buy something that has a too slow CPU that is not able to push enough traffic, but I also don't want to overkill and get a huge electricity bill.

Sebastian, How much CPU does your relay have/use?

From: tor-relays tor-relays-bounces@lists.torproject.org on behalf of Sebastian Hoffmann sebastian@machinewithoutbrain.de Sent: 17 March 2017 12:39 To: tor-relays@lists.torproject.org Subject: Re: [tor-relays] What kind of hardware do I need for my relay

Hi Farid, I'm running my Guard / Relay on a VPS with only one core and 512MB RAM. Its also hosting a blog, my own mailserver and is member of pool.ntp.org I haven't encountered any problems so far. It's also connected via 100MBit/sec You can see the stats here: https://atlas.torproject.org/#details/F381D6294A93A0078B76D0DCA332133CF5C8F6... https://atlas.torproject.org/#details/F381D6294A93A0078B76D0DCA332133CF5C8F687 I experimented a bit in the last weeks with being a bridge, but I now I'm back as a Relay... therefore the performance charts are not as high as they have been at the end of last year.

Am 2017-03-17 10:54, schrieb Farid Joubbi:

...

Hello,

I have been running low powered middle relays for a few years. Now I think it's time for a small upgrade. I haven't looked at computer hardware for a long time. What is the cheapest and most power efficient CPU/motherboard that can saturate 100 Mbit/s as a Tor relay?

Thanks.

---

tor-relays mailing list tor-relays@lists.torproject.org mailto:tor-relays@lists.torproject.org https://lists.torproject.org/cgi-bin/mailman/listinfo/tor-relays https://lists.torproject.org/cgi-bin/mailman/listinfo/tor-relays_______________________________________________

tor-relays mailing list tor-relays@lists.torproject.org https://lists.torproject.org/cgi-bin/mailman/listinfo/tor-relays

Agreed you won't need much to max out a 100 Mbps connection.

I have a few relays in the ~ 200 Mbps ballpark running on single core VPS instances. The underlying hardware is based on Intel Xeon E5-2650L.

CPUs with AES-NI help. 2 cores may offer marginal improvement. More than 2 cores isn't going to help at all. Tor still isn't optimized for it.

If running on Linux or Unix there are a lot of optimizations to be done. For Linux, I'd start here: https://www.torservers.net/wiki/setup/server and look at the "High Bandwidth Tweaks" section.

On Mon, 20 Mar 2017 22:27:59 +0000, Olaf Grimm wrote: ...

My personal usage drives the internet line to full power, but Tor as my MIDDLE RELAY doesn't use the full internet line power.

Which is a good thing, by the way. Tor traffic is bursty, so when your tor node actually saturating the link it means that the latency goes up, which means poor user experience. (This is generally so in networks - full pipes are bad.)

- Andreas

I do mean Megabits. I have learned a long time ago that Tor traffic throughput can't be compared with ssh. Tor needs to sustain the traffic amount for a long period of time with many variables affecting the measured bandwidth.

Also I would prefer to run OpenBSD, which does not have support for AES-NI by design (last time I checked).

________________________________________ From: tor-relays tor-relays-bounces@lists.torproject.org on behalf of Andreas Krey a.krey@gmx.de Sent: 20 March 2017 22:33 To: tor-relays@lists.torproject.org Subject: Re: [tor-relays] What kind of hardware do I need for my relay

On Mon, 20 Mar 2017 21:03:57 +0000, Farid Joubbi wrote:

I have tried a Banana Pi Pro 1,2 GHz Allwinner A20 -> 10 Mbit/s max (debian)

You do mean Mbit/s and not Mbyte/s? Even my old raspi B (first gen) needs only 30% CPU to process 12MBit/s (ssh), and my bananas transfer data via scp at 6 MByte/s (also ssh).

- Andreas

-- "Totally trivial. Famous last words." From: Linus Torvalds <torvalds@*.org> Date: Fri, 22 Jan 2010 07:29:21 -0800 _______________________________________________ tor-relays mailing list tor-relays@lists.torproject.org https://lists.torproject.org/cgi-bin/mailman/listinfo/tor-relays

OK. I thought from the beginning that my relay running the Banana Pi would be capable of handling more traffic. I have asked about it before, and got some really good answers. I still can't completely explain why it does not handle more. You can read about it here: https://lists.torproject.org/pipermail/tor-relays/2016-September/010142.html

The relay is still up, you can see a link to atlas on that thread. It has become a bit unstable lately and stops responding for some unknown reason. I don't know why the amount of traffic about doubled in October.

I ran a relay on a Intel NUC on OpenBSD for a few months. The traffic through that relay was very very low compared with my expectations. I got an answer from this list that the OS is most probably the one to blame. It has been shown that OpenBSD has a lower throughput of Tor traffic compared with FreeBSD and Linux. I have not found an answer to why it is so. Still I think it's a good thing for Tor to have the diversity of different OSes running the network.

I don't understand why Atlas shows the amount of traffic in Bytes instead of bits. I just know that it is so and that it has confused more than one person in the past.

So in conclusion, I'm more confused now than I ever was before ;-)

________________________________________ From: tor-relays tor-relays-bounces@lists.torproject.org on behalf of Andreas Krey a.krey@gmx.de Sent: 21 March 2017 20:53 To: tor-relays@lists.torproject.org Subject: Re: [tor-relays] What kind of hardware do I need for my relay

On Mon, 20 Mar 2017 22:49:53 +0000, Farid Joubbi wrote:

I do mean Megabits. I have learned a long time ago that Tor traffic throughput can't be compared with ssh.

No, but it can be used to roughly judge what the hardware is capable of. It doesn't help to throw more hardware at a node when it just doesn't get more traffic from the network.

Last year one of my node's traffic increased about tenfold in the span of a month, without reason or rhyme as seen in https://blog.apk.li/2017/01/29/tor-relay-traffic-again.html

- Andreas

-- "Totally trivial. Famous last words." From: Linus Torvalds <torvalds@*.org> Date: Fri, 22 Jan 2010 07:29:21 -0800 _______________________________________________ tor-relays mailing list tor-relays@lists.torproject.org https://lists.torproject.org/cgi-bin/mailman/listinfo/tor-relays

Hi,

I added 0x3d006 and 0x3d007 on a Dedibox XC (https://www.online.net/fr/serveur-dedie#perso) just a week ago. They offer 1 GBps both directions unmetered (+2,5 Gbps to the internal network) for very little money on an Intel Atom C2750 (with AES-NI, cpubenchmark.net score: 3929, single-thread rating: 579) with 16GB of RAM and 1TB of HDD (or 250GB SSD if you manage to grab one of those). Those nodes are abviously still picking up speed but right now they're already doing 8.73 and 5.6MB/s after only 8 days with both Tor processes eating on avg. around 15% of CPU. At 100% load per instance that would give me approx. 2 x 45 MB/s when running on full speed - altogether ~720Mbps which would barely be filling the pipe.

So on this particular machine 45MB/s per core (with a 579 points single-thread rating) should add up to ~13 single-thread points being 1 MB/s. That may of course vary depending on your CPU architecture and network but it just may be something to base your calculations on. Some of my other nodes (0x3d001 and 0x3d002) are running on an E3-1245 v2 and they both do around 40MB/s, single-thread score for that CPU is 1996 which would equal a bit less than 50 points per 1 MB/s. But on those instances it's rather the pipe that's the limiting factor, I can't push much more than 80MB/s there any direction any time of the day. Well, check the stats for yourself and do your own calculations: https://atlas.torproject.org/#search/0x3d00 (0x3d003 is running on an E3-1230 v2 and 0x3d004 and 005 are an i7 6700K).

*tl;dr:* On average around 25~30 single-thread points for your CPU on cpubenchmark.net should give you +/- 1MB/s of sustainable Tor traffic.

---- Andy Weber andy@0x3d.lu

On 28.03.2017 12:54, Petrusko wrote:

Hey Farid, Have you found an interesting low cost hardware since this last message ? Sometimes I try to look for it, but there's a lot of little cards like RPi, Banana... sadly I think it has not enough CPU power to play with a lot of Tor traffic :s

On the torserver webpage, there's a command line to know if the cpu has AES-NI acceleration.

cat /proc/cpuinfo | grep aes

For fun, I've tried on a laptop with Intel Core 2 Duo... no result shown after this command. So sadly this laptop will not be enough strong to have fun with this kind of crypto... it's sad because it's not burning a lot of watts!

Farid Joubbi :

...

OK. I thought from the beginning that my relay running the Banana Pi would be capable of handling more traffic. I have asked about it before, and got some really good answers. I still can't completely explain why it does not handle more.

---

tor-relays mailing list tor-relays@lists.torproject.org https://lists.torproject.org/cgi-bin/mailman/listinfo/tor-relays

Ok, N3050 cpu count was my mistake. It are 2 cores. It is a hand-size mini pc (ZOTAC 323 nano). It act as a middle node.

Now checked: The cpu load is ~15% with 1184 connected relays! (home connection with 10MBit/s; actual modem usage: 5MBit/s upload average, 8MBit/s peak. Downstream equal upstream.

(Sorry for wrong values before with an upload of 1MBit/s. I had forgotten my line upgrade some months ago.)

Relay fingerprint: E9A31E7EFF7A062AA67DA601CC70605AACF9F385 Relay external address: 91.48.47.161 Total traffic (read/written): 260.18 GB/263.10 GB Connected relays (1184)

ATLAS says 200 kByte/s (graph) >> ~2MBit/s Where is the mistake? I expect a lot more. The wrong value is stable over weeks.

One of my external VPS server:

Incoming / outgoing ~30MBit/s with a 100MBit/s connection

ATLAS says 4MByte/s (graph)

ARM display 40% cpu load with 5% for ARM itself.

RAM 4GB, 1 virtual core, XEON, ?? GHz

My confusion: ATLAS display MByte/s, not MBit/s. Every time I traped in this. It is a technical agreement, that serial lines are rated in Bit/s, not Bytes/s. It would be nice to change this in ATLAS.

I hope now my values are clearer than before.

Olaf

On 21.03.2017 06:07, Roman Mamedov wrote:

On Mon, 20 Mar 2017 21:03:57 +0000 Farid Joubbi joubbi@kth.se wrote:

...

Intel NUC5CPYH Celeron N3050 1,6 GHz to 2,16 burst -> 5 Mbit/s max (OpenBSD)

This sounds wrong. VIA Nano 1.6GHz, a single core laptop CPU from 2011, can sustain about 40+40 Mbit (WITHOUT utilizing the crypto acceleration).

The Celeron N3050 has two cores and supports AES-NI, and if OpenBSD seriously doesn't (which I doubt), then swap out OpenBSD.

With two cores you could also run 2 instances of Tor for more efficient CPU use. I would expect both cores stay at sub-50% CPU use while the entire thing is capping out into the full 100+100 Mbit.

That said, when did you measure the speed, and are you aware of https://blog.torproject.org/blog/lifecycle-of-a-new-relay ?

Another possibility, maybe your ISP only provides "up to 100 Mbit" and is shaping certain types, patterns or amounts of traffic use.