[headline]

I have a 1700x in the mail. I could do a few tests.

[Powerlord]

From what I remember, both Windows and Linux let you restrict processes to specific cores. That's cores, as in more than one.

In Linux, you use the taskset command when starting the process:

taskset 0xC srcds_run ...

In Windows, you use the start command when starting the process:

start /affinity 0xC srcds -console ...

The hexadecimal is actually a hex representation of binary for the processors.

Core 1 = 0b00000001 = 20 = 0x01
Core 2 = 0b00000010 = 21 = 0x02
Core 3 = 0b00000100 = 22 = 0x04 
Core 4 = 0b00001000 = 23 = 0x08 
Core 5 = 0b00010000 = 24 = 0x10
Core 6 = 0b00100000 = 25 = 0x20
Core 7 = 0b01000000 = 26 = 0x40
Core 8 = 0b10000000 = 27 = 0x80

So, in the examples above 0x04 (Core 3) + 0x08 (Core 4) = 0x0C

[8guawong]

Originally Posted by Headline I have a 1700x in the mail. I could do a few tests.

looking foward to your test

would be awesome if you can test 64 player zombie escape server

[headline]

Originally Posted by 8guawong looking foward to your test would be awesome if you can test 64 player zombie escape server

I don't think I have the bandwith to support 64 actual players.


However I'm pretty sure I can use sv_stressbots.

[sneaK]

Originally Posted by Headline I don't think I have the bandwith to support 64 actual players. However I'm pretty sure I can use sv_stressbots.

To my knowledge bots actually use more CPU than actual players, so that should work out fine.

I'm actually curious too, not that I'm ready to jump to AMD, but this new architecture certainly proves to be intriguing!

[headline]

Originally Posted by blackhawk74 To my knowledge bots actually use more CPU than actual players, so that should work out fine. I'm actually curious too, not that I'm ready to jump to AMD, but this new architecture certainly proves to be intriguing!

I ended up getting an 1800x, but results between the two won't be significant if you OC the 1700X (which you should do anyway)...

It'll be a few weeks until my new rig is up and running because I'll be in Germany all next week

[Kxnrl]

i used i3 7350K @5.0Ghz instead of 7700K.
cuz in 1U server.

[Robin955]

Originally Posted by Neuro Toxin Talking windows: I would avoid binding srcds processes to a cpu (even binding to two virtual cpu's which belong to the same physical cpu when hyperthreaded). SRCDS usage spikes alot! This is pretty much one instance per core on a Intel-Xeon-E3-1270-v5. When a server instance spikes and hits the core's max usage you see sv/var spiking badly when playing in the game server. Also notice most of the time srcds instances are not using much cpu at all. Below are the same instances with roughly the same usage not bound to cpu cores. Notice the srcds instances don't spike as high while the cpu is being used more consistently over time generating a higher usage readout. By allowing the scheduler to take control it's able to effectively switch the load between cpu cores. This means the scheduler can "attempt to" avoid cores maxing out by switching the srcds instance onto a new core that has less usage when required. In game, you see this as a harmless sudden bump in sv / var like so: Each time a srcds process switch's physical cores, it's cpu cache may need to be transferred to another physical core. When these cache misses occur the process is held up for a moment while it's cache is transferred causing the harmless minor sv/var spike. Now say the scheduler cant switch a processes cpu core because it's bound. That single cpu spikes out causing yellow/red sv/var readouts in game making the server lag like fuck. The mention on cache is because the faster / better the cache architecture is on the cpu, the faster the cache misses transfer the server processes onto new cores. Therefor reducing the amount of time before the server starts to process again not giving as big spikes in sv/var. With this being said you still stick to one srcds instance per virtual core and work out your max usage by looking at how many players start to max a single core out. If for some reason, all the servers cpu spikes decide to universally align, the cpu can throttle from each instance using every core at the same time and the scheduler trying to switch the servers around when every core is full. If you want to host stable as fuck servers with really small sv/var spikes. You can bind you cores, you will really massively need to manage player counts and maps to ensure their respective cores dont max out. Generally you wont be able to host as many overall player slots. If you want big as fuck servers. Let the scheduler switch a servers core over to get a little extra emphhh out of it. If your host lots of small 5v5 servers, let the scheduler do the work as you will be able to host lots more servers overall.

are you running with SMT OFF?
im getting this even with addons folder disabled / stock server, ehm but the VAR M/S is alot higher on your pics and is that then related to the rates? can the var i see be dependent on my network connection?


i also disagree with your core changing process, im on windows here and i've assigned each server to each cpu 0-16 and they all run 1 server per cpu and even on higher priority the VAR is still 0.1 and higher, i also have spikes randomly that goes 1.5 - 2.0 due to ryzen's infinity fabric.
the cache you're talking about is the L1 Cache and not the L3 Cache as the L3 is equal to xeons w/e 2MB per core. (alot of ppl can get confused as when you look at the product it will say 16mb cache or 4+16 = 20mb cache for the ryzen 8c16 cpu's.

atleast i got my servers assigned, the improvements are not there, the var is 0.1-0.5 and then it has random spikes to 2.0 but i cant feel it but i see it so its concerning, thats why i came here, im gonna try with SMT Disabled so i have 8Cores with 16MB Cache (4mb cache per core) se if that works out. that should provide me with twice the L1 and L2 Cache per core aswell.
also im running 3200 Mhz CL16 ram.

PS for ryzen users, make sure you manually overclocked it or else you have to manually disable turbo boost so the servers runs stable.

and CPU usage is not spiking for me > 20 bots with 50-70 plugins on the server and the cpu usage is below 50%
there is improvement with disabling SMT to give each cpu twice the cache but it isn't enough to make up for the extra 8 servers you could been hosting or 7.

also a question, what linux to run for the best server performance?

i also assume that your pic with 0.015 var was not from a ryzen cpu?
guess im the first one to give you guys the live results of ryzen.

[Deathknife]

Originally Posted by Robin955 are you running with SMT OFF? im getting this even with addons folder disabled / stock server, ehm but the VAR M/S is alot higher on your pics and is that then related to the rates? can the var i see be dependent on my network connection? i also disagree with your core changing process, im on windows here and i've assigned each server to each cpu 0-16 and they all run 1 server per cpu and even on higher priority the VAR is still 0.1 and higher, i also have spikes randomly that goes 1.5 - 2.0 due to ryzen's infinity fabric. the cache you're talking about is the L1 Cache and not the L3 Cache as the L3 is equal to xeons w/e 2MB per core. (alot of ppl can get confused as when you look at the product it will say 16mb cache or 4+16 = 20mb cache for the ryzen 8c16 cpu's. atleast i got my servers assigned, the improvements are not there, the var is 0.1-0.5 and then it has random spikes to 2.0 but i cant feel it but i see it so its concerning, thats why i came here, im gonna try with SMT Disabled so i have 8Cores with 16MB Cache (4mb cache per core) se if that works out. that should provide me with twice the L1 and L2 Cache per core aswell. also im running 3200 Mhz CL16 ram. PS for ryzen users, make sure you manually overclocked it or else you have to manually disable turbo boost so the servers runs stable. and CPU usage is not spiking for me > 20 bots with 50-70 plugins on the server and the cpu usage is below 50% there is improvement with disabling SMT to give each cpu twice the cache but it isn't enough to make up for the extra 8 servers you could been hosting or 7. also a question, what linux to run for the best server performance? i also assume that your pic with 0.015 var was not from a ryzen cpu? guess im the first one to give you guys the live results of ryzen.

I think all his pictures were taken on a Intel-Xeon-E3-1270-v5. with SMT on (it's 4core 8thread CPU)

If you think the var spikes are because of infinity fabric, you could lock half your servers to 0-3 and other half to 4-7 cores.

What clock speed were you running at and did your ram run at 3200 or at lower speed? There were/is? some issues with running high speed ram, 2933 seems to be the norm rn though.

[Robin955]

Originally Posted by Deathknife I think all his pictures were taken on a Intel-Xeon-E3-1270-v5. with SMT on (it's 4core 8thread CPU) If you think the var spikes are because of infinity fabric, you could lock half your servers to 0-3 and other half to 4-7 cores. What clock speed were you running at and did your ram run at 3200 or at lower speed? There were/is? some issues with running high speed ram, 2933 seems to be the norm rn though.

well servers seems to run just like any intel 4770k in terms of the var but it has some random spikes every 5-10th second to 1.5-2 var even with sourcemod disabled and stock server even when there's no people on.
thats my main concern, and no i got 3200 CL16 Stable, already ram memtest and so on.
ryzen was running 3.8ghz stable and temps about 30c