Someone broke the tor-relay speed record?
https://globe.thecthulhu.com/#/relay/F528DED21EACD2E4E9301EC0AABD370EDCAD2C4... Someone just got 149.08 MB/s on a non-exit relay. This is amazing! Would you mind saying what kind of hardware you use for this? Ipredator used https://ipredator.se/guide/torserver to get to 101MB/s. So your setup should be even more extreme. -- PGP fingerprint: 8516 5FFC 011A 6465 D042 6AC1 D719 1F41 B7CE EDFF OTR fingerprint OFTC: FF6D1AC2 13F38B15 EDB21522 D3530A6F 6EE01996 OTR fingerprint freenode: 293FAFD3 1155A465 C5B77015 B65E8AEC 63A1AF21 Full PGP public key: https://justaguy.pw
Oh wait? This is only advertised bandwith and not the actual bandwith. maybe the actual bandwith will reach the advertised bandwith some day. This relay is only running for 3 days so.. On 12/31/2014 10:11 AM, Justaguy wrote:
-- PGP fingerprint: 8516 5FFC 011A 6465 D042 6AC1 D719 1F41 B7CE EDFF OTR fingerprint OFTC: FF6D1AC2 13F38B15 EDB21522 D3530A6F 6EE01996 OTR fingerprint freenode: 293FAFD3 1155A465 C5B77015 B65E8AEC 63A1AF21 Full PGP public key: https://justaguy.pw
On Wed, 31 Dec 2014 01:13:52 -0800, Justaguy <justaguy@riseup.net> wrote:
The advertised Tor bandwidth for the exit node that I control matches up well with the bandwidth graph provided by the ISP, so I believe it is fairly accurate. https://globe.torproject.org/#/relay/E1E1059D8C41FC48B823C6F09348EA89C4D4C9D... Seems like it should be impossible however for a relay to jump to 149MB/s of advertised bandwidth in less than a week.
Seth schreef op 31/12/14 om 21:09:
The advertised bandwidth number is determined by the OR itself, not the directory authorities. It takes a simple change to the code to report higher numbers. The same applies to the bandwidth graphs in globe, those too come from the relay itself. Tom
A flag on the AS number where the ip of that Ipredator exit is in, does not mean that it is in this country. You can set that flag to anything you want. It's probably a joke about "libre" . On 12/31/2014 10:15 AM, I wrote:
-- PGP fingerprint: 8516 5FFC 011A 6465 D042 6AC1 D719 1F41 B7CE EDFF OTR fingerprint OFTC: FF6D1AC2 13F38B15 EDB21522 D3530A6F 6EE01996 OTR fingerprint freenode: 293FAFD3 1155A465 C5B77015 B65E8AEC 63A1AF21 Full PGP public key: https://justaguy.pw
On Wed, Dec 31, 2014 at 4:11 AM, Justaguy <justaguy@riseup.net> wrote:
Not within any definition of sane expenditure... The i7-4790k is Intel's highest clocked processor, ever. It's also Intel's current highest single thread performer. i7-4790k 4 x 4.0ghz cores = 1600 8mb 32gb ddr3-1600 88w $340 tsx-ni (1960 as OC'd by IPredator.) IPredator put a 10gbps nic in there and is moving only 800mbps OC'd (or 655mbps stock. Which is 165mbps/core, or 220mbps/core if isolcpus=1,2,3). Assume for moment this is cpu saturation. To do better you have to buy more core x ghz results like: i7-5930k 6 x 3.5ghz cores = 2100 15mb 64gb ddr4-2133 140w $580 So that's 1155mbps using all 6. Plus the more cache and much faster ram. And with good tor and os optimizations, you could probably hit 1250mbps. i7-5820k 6 x 3.3ghz cores = 1980 15mb 64gb ddr4-2133 140w $390 So that's 1090mbps plus. (ram for the above two: $420 32gb 4x8 ddr4-2133, non-OC) e5-2697v3 14 x 2.6ghz cores = 3640 35mb 768gb ddr4-2133 145w $2800 would fill 2000mbps plus. If you only need to fill 1000mbps, or however much bandwidth under that you're buying, just find the cheapest core x ghz that does that. While IPredator and their build may be "cool", it fails to ROI, thus it's completely pointless to build. They wasted... https://www.google.com/finance?q=EURUSD Capital cost: - Water cooling to get 22.5% clock over stock. (More common/free OC gains are the low end of 5-18%.) You could simply pay $240 more for the i7-5930k and get more core x ghz performance than they did, without overclocking. (Overclocking is also by definition a risk to data/lifetime). Or pay $50 more and still get more. (Both assuming ram and other resources aren't used up due to two more core of tor instances). $1110 radiator (IPredator quoted $1570) $55 flow meter (IPredator qooted $50) $30 connectors, estimated qty 2 x $15 $100 coolant, qty 5 x $20 (IPredator quoted $120) $25 tubing coil $5 clamps, estimated qty 10 $95 waterblock $15 paste (IPredator quoted $25) wasted: $1435 low to $1920 high - Ram $1000 32gb 4x8 ddr3-2800, equivalent brand (IPredator quoted $1550) $280 32gb 4x8 ddr3-1600, non-OC wasted: $720 low to $1270 high - Case $730 5u case (IPredator quoted $660, saved $70) $300 1u case wasted: $360 low to $430 high - Motherboard $25 estimated economy vs. enthusiast savings wasted: $25 total capital wasted: $2540 low to $3645 high net performance gain vs. other capital options: zero or less $760 10G-PCIE2-8C2-2S (IPredator quoted $660, saved $100) $440 E10G42BTDA is it comparable? if so that's another $320 wasted (Total cost IPredator quoted: $5420) Ongoing rent cost: - 4u on the chassis (1u required + 4u wasted) - 3u on the radiator wasted: 7u Enthusias[m|t] is great, Tor could use some, and I've no want to diminish that... but back in the real world away from the show... if you just ran the i7-5930k stock in a barebones 1U, you'd have enough capital left to buy 2 more servers and enough rent to feed them at least some internet. (IPredator failed to quote in their review the amount of bandwidth purchased, the price/megabit, and thier cpu load at a given megabit. This makes it harder to fully analyze. Please post cpu load @ mbps data.) Moral: Operators... please don't waste your Tor budget on silly enthusiast crap. I'd also suggest testing FreeBSD. If you want to run real Tor contests, try for... - most bandwidth per node cost (depth, as above) - most usable nodes per cost (breadth) - etc And if you want me to win them, or this saved you $$$, donate here :-) btc 1BbEqMvEdsKiPuRT75HGrjZP8zqquamBPn
Since people aren't going to like paying the 10g switchport fee, nor the price of small bandwidth over 1gbps on it, the fastest real world box for individual tor nodes is probably going to be that i7-5820k off a gig port for $1235 or less.
I would like to give a worthy mention about price/performance of modern AMD CPUs. My current highest observed peak on a $400 AMD (w/o rackcase) machine is 370Mbps running one tor process with two threads. On 2 January 2015 at 14:59, grarpamp <grarpamp@gmail.com> wrote:
On Sat, Jan 3, 2015 at 3:53 AM, usprey <usprey@gmail.com> wrote:
What cpu model and cpu load percentages at what bandwidth? The AMD's seem nice if you don't pay for watts, want ATI opencl GPU performance on die, can utilize their cmt setup better than htt, or just want a decent cheap cpu (which could indeed be interesting to evaluate for lower bandwidth like this 370Mbps). AMD's fastest processor seems the fx-9590 4 modules 4.7ghz 220w $230. Under full load, my guess is the i7-5820k is about 26% faster for $155 more, which you'd make back saving watts in 2.25 years. I don't know how much Tor benefits from SMT, or how Intel HTT vs AMD CMT compares (very few reviews bother to properly test SMT).
On 3 January 2015 at 12:36, grarpamp <grarpamp@gmail.com> wrote:
A8-5600K bought for a cheap private server 1,5 years ago https://en.wikipedia.org/wiki/List_of_AMD_accelerated_processing_unit_microp... I can set up system stats if you want proper figures, but far from full load with 250Mbps average bidirectional traffic.
I currently have free hosting for this machine, so power and traffic not an issue. I expect the kernel to handle the rest. ;)
Good point, if your not on a budget and building a private relay. Power prices also vary a lot between countries and regions.
I don't know how much Tor benefits from SMT, or how Intel HTT vs AMD CMT compares (very few reviews bother to properly test SMT).
No idea, should this not be more dependant on the OS?
On Sat, Jan 3, 2015 at 9:20 AM, usprey <usprey@gmail.com> wrote:
A8-5600K bought for a cheap private server 1,5 years ago https://en.wikipedia.org/wiki/List_of_AMD_accelerated_processing_unit_microp...
That could probably deliver 1/3 of an i7-5820k at 1/4 the price... a better deal if targeting its lower possible bandwidth.
I can set up system stats if you want proper figures, but far from full load with 250Mbps average bidirectional traffic.
Relative performance is mostly a benchmark thing, so the model lineup I gave can be found anywhere and is pretty solid. To be able to better predict/pick the megabits a model could deliver I'd need a handful of better report samples/graphs: - cpu model - graph of total system load % vs each cpu load % vs bandwidth - test environment statement, how many tors running, SMT, other system uses, bandwidth/nic config, OS... Most reports seen on this list are useless to establish that... "oh, i'm running an i3 and moving 60Mbps blah blah..." so don't take my megabit estimates as anything yet.
participants (6)
-
grarpamp -
I -
Justaguy -
Seth -
Tom van der Woerdt -
usprey