I run the non-exit relay: https://metrics.torproject.org/rs.html#details/855BC2DABE24C861CD887DB9B2E95... The relay run on a debian stretch machine with an i5-4670 at 3.8GHz with 4GB memory. CPU usage at 250Mbps traffic is around 40% of 1 core out of 4.
On April 1st my ISP doubled my bandwidth, from 250Mbps to 500Mbps. So far the Tor bandwidth authorities seems to not have picked up on all the new bandwidth. The observed bandwidth number has changed twice, increasing with small amounts.
How long does it take for the BW authorities to eventually observe a BW closer to 500Mbps. Weeks? Months? The reason I ask is that I wonder if I should run a second Tor instance or if the current one will be able to make use a a reasonable part of the 500Mps.
Am 06.04.2019 21:19, schrieb Logforme:
The reason I ask is that I wonder if I should run a second Tor instance or if the current one will be able to make use a a reasonable part of the 500Mps.
I'm also testing it with one to three instances. My problem is, I only have 30TB traffic / month. Unfortunately, that does not make sense with multiple instances. :-(
From https://www.torservers.net/wiki/setup/server:
Currently, Tor does not scale on multicore CPUs. If the CPU supports AES-NI crypto extensions (most modern CPUs do), one Tor process is able to handle around 400 Mbps of throughput – without AES-NI, around 100 Mbps.
Hello,
If Tor doesn't scale on multicore CPUs, setting NumCPUs to 2 and running two threads has no effect at all on throughput?
Thanks,
Conrad
On Sun, Apr 7, 2019 at 7:02 AM lists@for-privacy.net wrote:
Am 06.04.2019 21:19, schrieb Logforme:
The reason I ask is that I wonder if I should run a second Tor instance or if the current one will be able to make use a a reasonable part of the 500Mps.
I'm also testing it with one to three instances. My problem is, I only have 30TB traffic / month. Unfortunately, that does not make sense with multiple instances. :-(
From https://www.torservers.net/wiki/setup/server:
Currently, Tor does not scale on multicore CPUs. If the CPU supports AES-NI crypto extensions (most modern CPUs do), one Tor process is able to handle around 400 Mbps of throughput – without AES-NI, around 100 Mbps.
-- Ciao Marco! _______________________________________________ tor-relays mailing list tor-relays@lists.torproject.org https://lists.torproject.org/cgi-bin/mailman/listinfo/tor-relays
On 8 Apr 2019, at 07:57, teor teor@riseup.net wrote:
The reason I ask is that I wonder if I should run a second Tor instance or if the current one will be able to make use a a reasonable part of the 500Mps.
It looks like your relay could be CPU-core-limited, or limited by some other local resource, or limited by its location.
To work out where the limit is, run another Tor instance.
On 8 Apr 2019, at 13:00, Conrad Rockenhaus conrad@rockenhaus.com wrote:
If Tor doesn't scale on multicore CPUs, setting NumCPUs to 2 and running two threads has no effect at all on throughput?
On most systems, Tor automatically sets NumCPUs to the number of physical CPU cores.
Tor is partially multithreaded, so those extra threads do make a difference. But if your tor process is limited by main thread operations, then you need to run another tor process.
T
-- teor ----------------------------------------------------------------------
Am 08.04.2019 06:38, teor wrote:
On 8 Apr 2019, at 13:00, Conrad Rockenhaus conrad@rockenhaus.com wrote:
If Tor doesn't scale on multicore CPUs, setting NumCPUs to 2 and running two threads has no effect at all on throughput?
On most systems, Tor automatically sets NumCPUs to the number of physical CPU cores.
Yeah, I tested it once. (Intel XEON E3-1270 v2 + 32GB RAM) NumCPUs not set: tor runs with 8 threads. With NumCPUs=2: only with 2 threads.
On Wed, 10 Apr 2019 22:14:43 +0200 lists@for-privacy.net wrote:
Yeah, I tested it once. (Intel XEON E3-1270 v2 + 32GB RAM) NumCPUs not set: tor runs with 8 threads. With NumCPUs=2: only with 2 threads.
It does, but it still isn't able to split the load into 8 threads evenly at the moment. In my experience one Tor process will use at most 1.2-1.3 of a core. Though I didn't try running multi-gigabit relays, only 100 Mbit ones on slow (but multi-core) CPUs.
Hi,
On 7 Apr 2019, at 05:19, Logforme m7527@abc.se wrote:
I run the non-exit relay: https://metrics.torproject.org/rs.html#details/855BC2DABE24C861CD887DB9B2E95... The relay run on a debian stretch machine with an i5-4670 at 3.8GHz with 4GB memory. CPU usage at 250Mbps traffic is around 40% of 1 core out of 4.
On April 1st my ISP doubled my bandwidth, from 250Mbps to 500Mbps. So far the Tor bandwidth authorities seems to not have picked up on all the new bandwidth. The observed bandwidth number has changed twice, increasing with small amounts.
How long does it take for the BW authorities to eventually observe a BW closer to 500Mbps. Weeks? Months?
Your relay observes its own bandwidth, and tells the bandwidth authorities the maximum over the last 5 days.
Looking at the 6 months graph from 1 April, your relay's observed bandwidth has increased about 5-10%. A small increase per week isn't bad for a guard: even if your consensus weight goes up, it takes time for clients to rotate guards.
The bandwidth authorities also measure the excess bandwidth on your relay every few days, and combine their measurements with your relay's observed bandwidth to generate their consensus weight votes. The consensus value is the low-median of those votes.
Looking at the consensus weight graph, the votes haven't changed much at all.
(The consensus weight changes the number of clients that use your relay, which increases its observed bandwidth, but decreases the measured bandwidth. Eventually these changes balance out.)
The reason I ask is that I wonder if I should run a second Tor instance or if the current one will be able to make use a a reasonable part of the 500Mps.
It looks like your relay could be CPU-core-limited, or limited by some other local resource, or limited by its location.
To work out where the limit is, run another Tor instance.
You could also wait another week to see if your relay picks up another 5-10% traffic increase.
T
One more thing:
On 8 Apr 2019, at 07:57, teor teor@riseup.net wrote:
Hi,
On 7 Apr 2019, at 05:19, Logforme m7527@abc.se wrote:
I run the non-exit relay: https://metrics.torproject.org/rs.html#details/855BC2DABE24C861CD887DB9B2E95... The relay run on a debian stretch machine with an i5-4670 at 3.8GHz with 4GB memory. CPU usage at 250Mbps traffic is around 40% of 1 core out of 4.
On April 1st my ISP doubled my bandwidth, from 250Mbps to 500Mbps. So far the Tor bandwidth authorities seems to not have picked up on all the new bandwidth. The observed bandwidth number has changed twice, increasing with small amounts.
How long does it take for the BW authorities to eventually observe a BW closer to 500Mbps. Weeks? Months?
Your relay observes its own bandwidth, and tells the bandwidth authorities the maximum over the last 5 days.
Looking at the 6 months graph from 1 April, your relay's observed bandwidth has increased about 5-10%. A small increase per week isn't bad for a guard: even if your consensus weight goes up, it takes time for clients to rotate guards.
The bandwidth authorities also measure the excess bandwidth on your relay every few days, and combine their measurements with your relay's observed bandwidth to generate their consensus weight votes. The consensus value is the low-median of those votes.
Looking at the consensus weight graph, the votes haven't changed much at all.
(The consensus weight changes the number of clients that use your relay, which increases its observed bandwidth, but decreases the measured bandwidth. Eventually these changes balance out.)
The reason I ask is that I wonder if I should run a second Tor instance or if the current one will be able to make use a a reasonable part of the 500Mps.
It looks like your relay could be CPU-core-limited, or limited by some other local resource, or limited by its location.
It's probably a local resource, because the bandwidth authority measurements don't vary much, even though the bandwidth authorities are on two different continents: https://consensus-health.torproject.org/consensus-health-2019-04-07-20-00.ht...
To work out where the limit is, run another Tor instance.
You could also wait another week to see if your relay picks up another 5-10% traffic increase.
T
On 2019-04-07 23:57:58, "teor" teor@riseup.net wrote:
It looks like your relay could be CPU-core-limited, or limited by some other local resource, or limited by its location.
Currently the CPU is only using 40% of 1 core (out of 4). All of it from Tor when BW is close to 250Mbps When routing from the LAN (iperf, downloading stuff etc) the machine easily pushes 500Mbps. It can of course be other local resources that Tor needs. I have set a lot of networking parameters in sysctl.conf many years ago (from torservers,net I believe), but I don't really know what all of that stuff means. As to limit by location (Sweden), not much to do about that :)
To work out where the limit is, run another Tor instance.
You could also wait another week to see if your relay picks up another 5-10% traffic increase.
I'll wait a bit longer to see what happens and then run a second instance.
Thanks for an excellent reply
tor-relays@lists.torproject.org
-
Conrad Rockenhaus -
lists＠for-privacy.net -
Logforme -
Roman Mamedov -
teor