Benchmark your MT with /emergeblocks

[Fixer]

Decided to do very little and simple benchmark, to do it:
1. run clean minetest without config/cache
2. create world: seed: 1, mapgen v7
3. start it
4. look up on game start, press t and type: /emergeblocks (0,0,0) (-1000,-1000,-1000)
5. wait around 20 min... don't do anything at this time
6. at the end you will get "Finished emerging 262144 blocks in X msec (on screen/in chat console)
7. post your results with hardware specs, os, minetest version

My results:

Code: Select all

262144 blocks in 1219303.0 ms (or 1219 sec)
Core i3-2120 3.3GHz, 8Gb DDRam 2, wd green 1Tb 5400rpm SATAIII, AMD/ATI Radeon HD 6870, Win 7 sp1, Minetest 0.4.14-dev

[burli]

hmmmm

Code: Select all

minetest: /build/minetest-9s30Wm/minetest-201606220316/src/emerge.cpp:456: bool EmergeManager::popBlockEmergeData(v3s16, BlockEmergeData*): Zusicherung »count_peer != 0« nicht erfüllt.
Abgebrochen

This error was from the daily build from the Ubuntu PPA. My self compiled version run's without any problems.

But this "benchmark" is pretty useless because it only benches one part of Minetest. But it is interesting to see that not all areas are generated with the same speed. It took a long time e.g. from 40% to 42%, but from 42% to 48% it was fast. Would like to know which areas took so long.

And it is not consistant, however. First I run with sqlite3, which needs 1250 seconds. Then I replaced the backend in world.mt with dummy. This time it tooks 1650 seconds. Then I try leveldb which also needs over 1600 seconds. Just to confirm my first result I set the backend to sqlite again, delete the db file and run again. Over 1600 seconds.

So what does that mean? I don't know

[Fixer]

Ha, this is just an experiment :)

[burli]

Ha, this is just an experiment :)

How dumb we are? ;-)

No, it is really an interesting experiment, but not a competition

[DI3HARD139]

Code: Select all

Finished emerging 262144 blocks in 946085.00ms .

(946.085 secs)
Intel Xeon X5670 @3.3GHz, 20GB DDR3 @1333 w/ 20GB Page File, 320GB WD Blue Sata II 3GB/s, GTX660 3GB, Windows 10 Pro x64, Minetest 0.4.14-release.
------------------------------------------------------------------------------------------------------------------------------------------------------------------
Edit:

Code: Select all

Finished emerging 262144 blocks in 655039.00ms . 

once the video I had encoding in the background finished.

[Fixer]

Since I have Intel CPU and updated OS due to meltdown fix, I decided to rerun this benchmark on 0.4.14-dev,
I've turned off av and io related stuff.

Result: 1388,711 sec, that is 13.9% slower.

We need more benchmarks to see how spectre/meltdown fixes affect minetest(especially server) performance.

P.S. 0.5.0-dev gives me just 706 sec on average.

[philipbenr]

262144 blocks in 257585.00ms (257.6 seconds, or 4:28 if I did that math right)

Specs:

• AMD Ryzen 7 1700 ( @3.4GHz )
• 2x8GB G.Skill Trident Z RAM ( @3.2GHz, 14-14-14-32 )
• XFX Radeon RX 580 8GB ( @1.41GHz )
• OCZ Trion 150 256GB SSD
• Windows 10 Pro
• 0.4.16

I did have a large installation going on in the background, as well as youtube, but with the 16 threads I have, I think the performance hit was pretty negligible.

[torbato]

I know it's an old thread but I needed something to do in this boring afternoon and it seemed like a fun topic lol

Here is mine:

Finished emerging 262144 blocks in 177776.00ms.

Specs:
Intel Core i7-1165G7 (4c/8t, max 4.70GHz, 12MB L3 Cache)
Intel Xe Graphics 96Eus
32GB 3200Mhz RAM
1TB Samsung NVME SSD
Windows 11 Pro
Minetest 5.5.0

[Festus1965]

Ha, this is just an experiment :)

How dumb we are? ;-)

No, it is really an interesting experiment, but not a competition

if would be more interesting if on one same machine different versions might be compared, as maybe 5.1, 5.2 5.3 5.4.1 and 5.5 under other things keep same : is there an slower or faster engine now.

As mentioned at first post with no cache sounds like VPS similar, but also CPU, RAM (Type+size) and kind of disc (HDD, SSD, M.2) will do also some influence.

So far not really anything worth yet, as not possible to extract some helpful data for config of hardware or even database.

And what does /emergeblocks have to do with 'real mts usage' ? what is a main decision for server after maybe mod composition.

So /emergeblocks may make sense on the same hardware OS with different mt versions,
but you can't overgo the bonds between:
* hardware
* OS
and at least the mt as a package.
Measure only this is maybe a small part of mtserver, but if your really interested in fast mtserver - you need to check all.

[sorcerykid]

I would imagine that emerge blocks might face a bottleneck with disk I/O, so a ramdisk would probably perform the fastest during such a stress test, compared to a mechanical hard drive and even SSD.

[Festus1965]

I would imagine that emerge blocks might face a bottleneck with disk I/O, so a ramdisk would probably perform the fastest during such a stress test, compared to a mechanical hard drive and even SSD.

test it ?
* HDD 150 MB/s
* SSD 450-550 MB/s
* M.2 ~1.500-3.200 MB/s so far I see
* RAM write cache ~ 15.000 MB/s
it must not be a RAM Disk, just to be able to use the writing cache on the way to discs might do the job. As at my server.

as of another reason did it: (26.03.2022)
minetest.conf just "seed and mapgen' and name = x
new empty map via standard sqlite, but all settings at server same ... vm.swappiness, write cache ... as every day.

Code: Select all

... /262144 Blocks ...
in 247128.11 ms

or 247 sec

Code: Select all

I7-4770 3.4 GHz, 32 GB DDR3 RAM (1333 at) 1600 MHz, M.2 SSD 128 GB, max 2.800 rw or was it 3200, Ubuntu 20.04, mt 5.5.0

so I am faster then philipbenr by 10 s but behind torbato 177s.

So what will happen if I move the DB to postgresql ?
and then also do the test two times, force that this area keep in cache ?

looking at htop during this time, only ONE thread sometime goes near 100% - CPU main is not hard working,
so that faster CPU from above testing are not the reason,
but both have double fast RAM 1600 against 3200,
as I use M.2 disc with 2.800/3.200 rw also hmm
as of only one thread as of standard config, or as of bottleneck at IOs?

using maybe cache, doing the test a second time just again after 10 minutes:

Code: Select all

... 26657.37 ms

or 27 sec ... just to know what cache is worth about !!
what your mt show with /status after the test ? here max_lag 0.332 s

[Festus1965]

nicely admin of Your-Land.de
did this test on his home test server ...

better CPU, double of mordern RAM, bigger M.2

and was slower (444s) in 5.5.0 than me, but at second time much faster (2.8 s)
guess MT internal cache did help then

in combination of also do the VPS IO tests,
with IO write/read without cache his PC is faster, (better CPU, faster M.2)
but as soon it uses cache my one is much faster.

so: must be a setting in OS to allow Cache overall ... also we check now fstab.

[bosapara]

Let's continue benchmarks.

Code: Select all

MT 0.4.17.1 - 274478 ms; MT 5.1.0 - 232636; MT 5.7.0 - 177820 ms

Code: Select all

i7 8700, 3.2ghz, 32gb 2600mhz, 240gb nvme, Ubuntu 20

Due to the full single-threaded operation within full test, I suppose the processor with a higher frequency will complete the benchmark faster.

[Walker]

Code: Select all

MT 5.6.0 - 142906.06ms

Code: Select all

AMD Ryzen 5950x, non-OC, 3,4GHz, 128GB DDR4 3200MTs, 2TB Samsung 980PRO

Code: Select all

Ubuntu 22.04 LTS x64 / Linux Kernel 5.15.0-47