A simple comparison of performance optimizations for gLUA
These are a few benchmarks that are tested on a Garry's Mod Server idling with DarkRP with only 1 player online. Some of those benchmarks, if possible, were also tested in a LUA5.3 console on Debian 10. If not otherwise stated differently all tests have been done on a 32bit gmod server on linux.
These benchmark results mean nothing compared to other benchmark results. Your numbers will probably look completely different than my numbers. Please compare your benchmark results only with your own numbers, as every small change in Hardware changes the result notably.
The goal of this collection of benchmarks is to show the actual performance gain by implementing "Coding Tips" from the web.
Everyone can test the code for themselves and see the difference. Source code is always linked at the bottom of a comparison.
The sources for some of the "Coding Tips" are:
lua-users.org
lua.org
stackoverflow.com
If you want to read more about gmod lua then you could also visit my wiki at luctus.at/wiki
Garry's Mod doesn't use only lua. It uses lua-jit.
This means the following benchmarks show, where applicable, also a comparison of lua and lua-jit.
Many of the performance tips online are supposed to be for lua, not luajit.
An example of lua vs lua-jit with the "multiplication vs division" example:
$ time luajit-2.1.0-beta3 mult_vs_div.lua
Multiplication: 0.00111624
Division: 0.00110691
real 0m1.255s
user 0m1.054s
sys 0m0.182s
$ time lua5.3 mult_vs_div.lua
Multiplication: 0.0284177
Division: 0.03088507
real 0m21.773s
user 0m21.623s
sys 0m0.120s
If you find any error, mistake or completely different result on your side please create a pull request.
I want this page to be as accurate as possible so that everyone can see and learn what actually benefits your code's performance.
TL;DR: If you use a a function, e.g. LocalPlayer(), often in a HUDPaint hook then I highly suggest caching it once at the beginning.
This is a comparison of the following code bits. The question is: "Should you cache a function you only use 3 times?".
The unoptimized version:
draw.RoundedBox(0, LocalPlayer():Health(), 500, 500, 500, color_white )
draw.RoundedBox(0, LocalPlayer():Health(), 200, 200, 200, color_white )
draw.RoundedBox(0, LocalPlayer():Health(), 10, 10, 10, color_white )and the better version:
local lph = LocalPlayer():Health()
draw.RoundedBox(0, lph, 500, 500, 500, color_white )
draw.RoundedBox(0, lph, 200, 200, 200, color_white )
draw.RoundedBox(0, lph, 10, 10, 10, color_white )The result (10000 frametimes):
Non-Local: 0.00001756282000116 (1.756282000116e-05)
Local: 0.000006777780000516 (6.777780000516e-06)
Code: files/hudpaint_3call_cache.lua
TL;DR: It is faster to cache the response of LocalPlayer() and overwrite the function to always return this value.
Pretty self explanatory. LocalPlayer() always returns you, the player, but takes more time than if you were to simply cache the player entity and return it everytime.
Check the code for more info.
Result:
--- Benchmark complete
On Client
reps 20 rounds 50000
old LocalPlayer 7.4946100011402e-08
fastLocalPlayer 6.2921299662776e-08
The Code: files/localplayer_cache.lua
TL;DR: It is a bit faster to "ply = ply or LocalPlayer()" instead of always using LocalPlayer(), but the time save is barely noticable.
Pretty self explanatory. This caches LocalPlayer() on first run and then always reuses this value.
Check the code for more info.
Result:
--- Benchmark complete
On Client
reps 20 rounds 10000
Cached 5.8824900024945e-07
Plain 6.021605001348e-07
The Code: files/localplayer_cached_or.lua
TL;DR: It is a faster to reuse the same Angle object and zero it before use than to create a new Angle object everytime.
The wiki states this as fact, and it is true. It saves you the overhead of creating a new object and garbage-collecting the old one.
Check the code for more info.
Result:
--- Benchmark complete
On Server
reps 30 rounds 5000
Zero'd 1.3485066712595e-07
Create 2.3365133318293e-07
--- Benchmark complete
On Client
reps 30 rounds 5000
Zero'd 1.1277866673481e-07
Create 2.4979333345679e-07
The Code: files/angle_zero_vs_new.lua
TL;DR: It is faster to cache the response of ply:SteamID() and overwrite the function to always return this value for a player from a caching table.
Pretty self explanatory. Same as "Caching LocalPlayer function" above.
Result:
--- Benchmark complete
On Server
reps 20 rounds 50000
SteamID Classic 4.6951399998613e-07
SteamID Cached 7.0017300033214e-08
--- Benchmark complete
On Client
reps 20 rounds 50000
SteamID Classic 5.2189459993372e-07
SteamID Cached 6.247109999822e-08
The Code: files/steamid_cache.lua
TL;DR: getDarkRPVar is always faster on server while on client it seems to be random. I would recommend using getDarkRPVar, as it is more optimized for network usage in comparison to NW1's constant network traffic every 10 seconds.
This is a comparison between ply:GetNWString("job") and ply:getDarkRPVar("job").
Result:
--- Benchmark complete
reps 10 rounds 10000
On Server
darkrpvar 1.2346400001078e-07
gmodnwvar 2.0404500001689e-07
--- Benchmark complete
reps 10 rounds 10000
On Client
darkrpvar 2.4343500003823e-07
gmodnwvar 2.036369999766e-07
As you can see, the result is different on client and server. The result on the client is always different everytime i run these tests. Sometimes, on the client , the gmodnwvar is faster by 25% and other times it is 300% slower than darkrpvar.
Code: files/nwvar_vs_darkrpvar.lua
TL;DR: Using a table is slower than using a variable.
This is a comparison between using a config table like myaddon.config.color = and a simple variable like myaddon_config_color =. This mainly exists because some people don't believe me.
The result:
--SERVER
tab= 1.9999993128295e-07
var= 1.0000007932831e-07
--CLIENT
tab= 1.0000007932831e-07
var= 0
As you can see above, using a single variable instead of a table structure is faster.
Getting a variable is as simple as "getting" it. But with a table you have to get the table and then the table element inside.
You can inspect this with lua's luac command, e.g. luac5.4 -l file.lua.
An example: print(abc) vs print(a.b.c)
$ luac5.4 -l var.lua
[...]
GETTABUP 0 0 2 ; _ENV "print"
GETTABUP 1 0 0 ; _ENV "abc"
CALL 0 2 1 ; 1 in 0 out
RETURN 0 1 1 ; 0 out
$ luac5.4 -l tab.lua
[...]
GETTABUP 0 0 4 ; _ENV "print"
GETTABUP 1 0 0 ; _ENV "a"
GETFIELD 1 1 1 ; "b"
GETFIELD 1 1 2 ; "c"
CALL 0 2 1 ; 1 in 0 out
RETURN 0 1 1 ; 0 out
As you can observe above, the table variant simply has more executions and is thus slower in theory and practice as shown above.
Code: files/table_vs_variable.lua
TL;DR: They are not as different as people make it out to be. Pairs is slower because it also works on non-sequential tables.
The result (10 tests of 100 runs each with calculating the sum of a 10.000 number table):
pairs 8.8780000373845e-07
ipairs 2.4300000836774e-07
for #t 1.639000013256e-07
Code: files/pairs_vs_ipairs_vs_for.lua
TL;DR: The fastest is the Iterator, but I (and the wiki) would not recommend using it as it is error prone. Using ipairs is a very readable, easy to use and fast way to iterate over all players.
The result:
pairs 2.7375729999562e-06
ipairs 2.5955539999833e-06
for #t 2.5866950000025e-06
iter 1.7860300003008e-07
Code: files/player_getall_compare.lua
TL;DR: Numbered and Sequential tables are always faster thanks to ipairs and for loops being able to speed up the loop. There is no difference between the string index length in terms of speed. Sequentially-numbered tables are faster than non-sequential number tables.
With a 100 element table and 1000 runs for each of the 5 testruns:
short string pairs 1.016059995527e-06
big string pairs 9.8490000327729e-07
sequent int pairs 8.060999940426e-07
sequent int ipairs 2.9101999789418e-07
unsorted int pairs 1.0006600032284e-06
There is barely any difference between long and short string tables. The main difference in terms of speed comes with sequential number index tables, which can utilize ipairs to be faster than using pairs.
Code: files/string_vs_number_table.lua
TL;DR: The time difference between Distance and DistToSqr is barely noticable.
The result (average of 10000 calculations):
Distance: 0.00000026040000557259 (2.6040000557259e-07)
DistToSqr: 0.00000025130000176432 (2.5130000176432e-07)
The Code: files/distance_vs_disttosqr.lua
TL;DR: They take the same amount of time. The difference is randomly swinging a tiny bit to each side every test.
Tested: a = t and 7 or 1 vs a=1 if t then a=7 end
The result (average of 1000000 calculations):
if: 4.1537399886693e-08
ternary: 4.168440006697e-08
The Code: files/ternary_vs_if.lua
TL;DR: The time difference between using table.insert(myTable,9) and myTable[#myTable+1] = 9 is barely noticable, as the results were mostly the same. (Even if tested on Lua5.3 console)
The result (inserting 1million values):
table.insert: 0.066328400000003
local table.insert: 0.07047399999999
[#res+1]: 0.066876700000023
The Code: files/table.insert_vs_table.lua
TL;DR: Directly checking the table key with table[value] is way faster.
The result (1million lookups):
table.HasValue: 0.032319600000051
table[value]: 0.00024150000001555
The Code: files/table.hasvalue_vs_table.lua
Note:
This is a O(n) vs O(1) situation. This means the bigger the table the slower table.HasValue will get.
For other examples visit https://wiki.facepunch.com/gmod/table.HasValue
To learn more about the "Big O Notation" visit, for example, https://web.mit.edu/16.070/www/lecture/big_o.pdf
TL;DR: Assigning an empty table is overall faster than emptying it.
The following test compares small (10 elements) and big (10.000) tables with each method. It also differentiates between numerical tables (key 1,2,3,...) and string index tables (key a,b,c,...).
Result:
--- Benchmark complete
reps 1 rounds 1000
On Server
table.Empty numerical on small 1.5931000018554e-06
table.Empty numerical on big 5.295060000276e-05
table.Empty stringind on small 4.9099999841928e-07
table.Empty stringind on big 4.7529999847029e-07
table = {} numerical on small 9.8249999609834e-07
table = {} numerical on big 6.0400000597838e-07
table = {} stringind on small 5.8179999678032e-07
table = {} stringind on big 4.8149999975067e-07
The above times vary greatly by table count and amount of tests included. In general, assigning a new table is faster than emptying one, the only exception being small string-indexed tables. The biggest difference is between emptying a big numerical table and assigning a new one.
An advantage of always assigning a new table: The time taken is nearly constant, so you don't have to worry about it.
The Code: files/table_empty_vs_new.lua
TL;DR: Not using metatables is faster.
This is a comparison of using the MetaTable PLAYER vs simply passing the player as an argument.
The result (100.000 rounds, 100 times):
ply:IsVal() 2.5642945799856e-06
IsVal(ply) 1.4792759700133e-06
As you can see above, not using the metatable is faster. This is because with ply:Func() you have to check if the function exists (which is calling the __index method of the PLAYER metatable) which is slow, while in comparison with Func(ply) you already have the function at hand.
The code: files/meta_vs_argument.lua
TL;DR: It is faster to add strings together via tables instead of stringing them together directly.
The result (10000 strings):
GMOD SERVERSIDE CONSOLE
String Concat: 0.32185959999993
Table Concat: 0.0099451999999474
LUA5.3 CONSOLE
String Concat: 0.086629
Table Concat: 0.006428
The Code: files/string_vs_table_concat.lua
TL;DR: They both have nearly the same speed of execution. It varies between who is faster every run, it's that close.
What is better depends on the use case, I recommend using string.format for cases where the input is not fixed, because with this function you can add entities/players without having to convert them to a string explicitly.
This is a comparison of the 2 ways of creating a single string out of multiple pieces, either with string.format or with .. between the arguments. Example:
format = string.format("%s has steamid %s",ply:Nick(),ply:SteamID())
concat = ply:Nick().." has steamid "..ply:SteamID()Result:
--- Benchmark complete
reps 10 rounds 1000
On Server
concat str 8.088840000346e-06
format str 8.3178000002022e-06
concat int 8.2556100002705e-06
format int 7.6555600000233e-06
As you can see above, the speed is varying a lot and here the format function is tied with the string concat.
The Code: files/string_format_vs_concat.lua
TL;DR: Using draw.SimpleText is nearly as fast as using the surface.* functions. DrawText is a bit slower and SimpleTextOutlined is by far the slowest.
This test simply draws a word on 3 different coordinates on your screen with 4 different functions and measures the time taken.
Result:
surface 4.4340000022203e-05
simple 4.8489999971935e-05
drawtext 5.563999984588e-05
outlined 0.00012634444445009
As you can see above the surface.* method is the fastest closely followed by draw.SimpleText. It is (maybe) faster than draw.DrawText because it doesn't "handle newlines properly" according to the gmod wiki. SimpleTextOutlined is the slowest because it has to draw an extra outline in comparison to the 3 other functions.
The Code: files/surface_vs_draw_text.lua
TL;DR: Using file.Read is a tiny bit faster but there is barely any difference between the 2 methods.
If you want to save text and get it later, which method is faster: Saving it to a file and reading it or saving it into sqlite and reading it?
Result:
--- Benchmark complete
reps 20 rounds 100
On Server
file 0.00013605510000012
sql 0.00014516330000007
There is no real speed difference, so it is not recommended to pick your storage type based on speed. You should pick your storage type based on what you need and how often you want to change data.
The Code: files/file_vs_sql_text.lua
TL;DR: They are not comparable as Query is a C function and the others LUA functions using it.
If you print the "short source" of these 3 functions with debug.getinfo you get:
Query [C]
QueryRow lua/includes/util/sql.lua
QueryValue lua/includes/util/sql.lua
The (shortened) source of QueryRow is:
function sql.QueryRow(query,row)
row = row or 1
local r = sql.Query(query)
if r then return r[row] end
return r
end
This means that it doesn't really matter (for speed) if you use Query and select the first row or use QueryRow directly.
TL;DR: ents.FindInBox and ents.FindInCone are the fastest ways to find entities close to a player because they are easy to calculate and do not return too many entities.
The tested methods of getting entities around a player are:
- ents.FindInPVS
- ents.FindInCone
- ents.FindInSphere
- ents.FindInBox
Result:
--- Benchmark complete
On Server
reps 10 rounds 10000
inPVS 2.2498948000473e-05
inBox 1.9315759995857e-06
Sphere 3.8246590005724e-06
Cone 1.6260100000977e-06
The Code: files/finding_entities.lua
TL;DR: It is faster to loop through all players and check their distance to you instead of using ents.FindInSphere and filtering for players.
This was tested with 40 players (bots) on a server. By looping through all players and checking the distance between you and them you are faster than by finding and looping through all the entities near you and checking if they are a player.
Result:
--- Benchmark complete
On Server
reps 10 rounds 1000
FindInSphere 0.00011627801000012
PlyGetAllChk 1.9475780000045e-05
--- Benchmark complete
On Client
reps 10 rounds 1000
FindInSphere 8.301507999995e-05
PlyGetAllChk 2.4392489999832e-05
The Code: files/finding_players.lua
TL;DR: As expected, the more uniqueness you need the longer it takes to calculate it. This means MD5 is the fastest but has a higher probability of colissions than the slower SHA256.
CRC32 and Base64 are just in there as comparisons. They are not a "good hash algorithm" in the same category as e.g. MD5.
Every function tested is defined in C and not in LUA. (According to debug.getinfo's short_src)
Result:
--- Benchmark complete
On Server
reps 10 rounds 10000
CRC32 4.5670400035306e-07
MD5 2.5526380002202e-06
SHA1 3.5287520001384e-06
SHA256 7.7691459993503e-06
Base64 4.1977299973951e-07
The Code: files/hashes.lua
TL;DR: Using the distance between normalized vectors of players viewing direction is the fastest method.
The question is: How do you calculate "Do these 2 players see each other on their screen right now?" the fastest?
This was a problem for a gameplay mechanic which needed to be calculated efficiently to not cause any lag.
Result:
--- Benchmark complete
On Server
reps 5 rounds 100
AngleBetweenVectorsManual 1.1895999687113e-06
AngleBetweenYawOnly 6.7819999458152e-07
WikisLookingAtVecDotFunction 1.0634000118444e-06
DistanceBetweenNormVecManual 1.5873999782343e-06
DistanceBetweenNormVecAuto 1.0834000095201e-06
DistanceBetweenAimVectors 8.4879999485565e-07
The fastest way is to use the ply:GetAimVector() of both players, negate one of those vectors and check the distance between them.
The default problem that many know is: Do these 2 vectors point in the same direction.
This problem is similar to ours: Do these 2 players look at each other.
The main difference is: One of the vectors is reversed.
This means we just have to reverse one of the 2 vectors and then we can use the standard mathematical formulas to calculate the "default problem" I mentioned above.
The slowdown while calculating this comes from using math.acos (arc cos) in the official formula. You need this to calculate the vector from the vector dot product formula.
An easier way is to simply use the distance between the 2 AimVectors. GetAimVector returns a normalized Vector which always has a length of 1. This means the maximum distance between the 2 vectors is 2 (if the vectors look in exactly the opposite direction) and the minimum is 0 (if the vectors look directly at each other). This makes it faster than all the other methods because you skip an angle calculation and just compare the distance directly.
The Code: files/looking_at_each_other.lua
TL;DR: It is faster to call a hook once with all players looping inside than to loop through players and calling a hook on each one by one.
This is a comparison between:
- Calling a hook for every player
- Calling a hook with a list of every player
Result:
--- Benchmark complete
On Server
reps 10 rounds 10000
single 9.2695600087609e-07
many 2.602249987649e-07
The Code: files/hook_many_vs_hook_once.lua
TL;DR: DarkRP's fn library is (most of the time) slower and more difficult to understand than simple lua code.
This is a comparison between 2 functions that do the same but are written completely different.
Please look at the Code to see exactly what the 2 tested codepieces were.
Result:
--- Benchmark complete
On Server
reps 10 rounds 1000
darkrp 7.9022700008863e-06 (0.0000079022700008863)
default 8.2877000014605e-07 (0.00000082877000014605)
Not only is DarkRP's fn code more difficult to read but it is up to 10 times slower than a simple chain of functions that do the same.
The Code: files/darkrp_fn_vs_standard_lua.lua
TL;DR: It is very slightly faster to use surface.DrawRect instead of draw.RoundedBox.
It should not be important for performance though as draw.RoundedBox also supports round edges and is simpler to read and write.
surface.DrawRect: 7.4859800010245e-06
draw.RoundedBox: 8.335099999681e-06
The Code: files/surface_vs_draw_box.lua
TL;DR: It is faster to call SetDrawColor with the rgb colors split up than if you were to use a Color object.
You can use the SetDrawColor function in 2 ways:
- surface.SetDrawColor(color_white)
- surface.SetDrawColor(255,255,255)
This is a comparison between these 2 methods. "Full" means passing the color object and "Split" means passing 3 numbers (rgb) to the function.
--- Benchmark complete
On Client
reps 10 rounds 10000
SetColorFull 6.0781200035763e-07
SetColorSplit 1.1340800017479e-07
The Code: files/setdrawcolor_comparison.lua
TL;DR: It is faster to assign all your variables in a table in the opening brackets.
I see 2 different ways of initializing tables in code nowadays.
The first simply being
local tab = {}
tab.name = "Test"
tab.money = 9and the second being
local tab = {
name = "Test",
money = 9,
}This is a comparison of both:
--- Benchmark complete
On Server
reps 10 rounds 10000
table = { x = y, 4.5506299968565e-07
table.x = y 6.2055800030748e-07
As you can see, it is faster if you create the table with the data already inside.
The Code: files/table_assignments.lua
TL;DR: MsgN and Msg are faster than print. Msg is only a bit faster than MsgN, because it doesn't append a newline at the end.
The test simply prints messages by using Msg, MsgN and print and checks the time taken.
Result:
--- Benchmark complete
reps 90 rounds 10
On Server
Msg 1.9603333061645e-06
print 3.1246666549123e-06
MsgN 2.0377778006756e-06
Msg vars 2.9552222390016e-06
MsgN vars 2.9710000045371e-06
print vars 9.2549999974128e-06
MsgN vars+tab 3.3575555587757e-06
The test is split into 2 sections: The first one where we only print a single string and the second one where we print varargs of 2 strings and 1 number.
As you can see, print is always the slowest.
The last benchmark is using MsgN with tabs to immitate what the print function does, and it is still faster than using print directly.
The Code: files/print_vs_msgn.lua
TL;DR: In Garry's Mod it is not faster to calculate x * 0.5 than x / 2. BUT: It is faster to multiply instead of divide in LUA5.3.
The result (average of 100 rounds with 1000000 calculations each):
GMOD SERVERSIDE CONSOLE
Multiplication: 0.00081622600000003
Division: 0.00081775100000215
LUA5.3 CONSOLE
Multiplication: 0.01865942
Division: 0.02342666
The Code: files/multiplication_vs_division.lua
TL;DR: It is a very tiny bit faster to use math.min in combination with math.max instead of math.Clamp, but the difference is negligible.
Result:
--- Benchmark complete
On Server
reps 20 rounds 50000
clamp 6.9868600001882e-08
max/min 6.806609987666e-08
--- Benchmark complete
On Client
reps 20 rounds 50000
clamp 6.9836399990663e-08
max/min 6.7090800087044e-08
The Code: files/math_clamp.lua
TL;DR: It is not faster to overwrite and implement the vector dot calculation in LUA.
Result:
--- Benchmark complete
On Server
reps 20 rounds 30000
vec:Dot old 1.2875416667678e-07
vec:Dot new 5.2033999999855e-07
The Code: files/vector_dot_overwrite.lua
TL;DR: It is not faster to overwrite and implement the vector Normalize calculation in LUA.
Result:
--- Benchmark complete
On Server
reps 20 rounds 30000
vec:Normalize old 1.1955383335495e-07
vec:Normalize new 5.1291399999911e-07
The Code: files/vector_normalize_overwrite.lua
TL;DR: It is not faster to overwrite and implement the vector Length calculation in LUA.
Result:
--- Benchmark complete
On Server
reps 20 rounds 30000
vec:Length old 1.1400050000068e-07
vec:Length new 3.8188000000545e-07
The Code: files/vector_length_overwrite.lua
TL;DR: It is a tiny bit faster to overwrite and implement the math.Round calculation in LUA.
Result:
--- Benchmark complete
On Server
reps 20 rounds 30000
math.Round old 7.1794499987448e-08
math.Round new 6.0206499994801e-08
The Code: files/math_round_overwrite.lua
TL;DR: It is a little itsy-bitsy bit faster to make all variables local in Garry's Mod. (14% in this example)
The result (average of 100 rounds with 10000 calculations each):
GMOD SERVERSIDE CONSOLE
global math.sin: 0.0000059390000023996 (5.9390000023996e-06)
local math.sin: 0.0000056840000024749 (5.6840000024749e-06)
LUA5.3 CONSOLE
global math.sin: 0.00057612
local math.sin: 0.00049858
The Code: files/local_vs_global.lua
TL;DR: It is not faster to cache ConVars in LUA.
The wiki page states that the GetConVar function "... caches the result in Lua for quicker lookups.", so caching the result again won't make it that faster.
Result:
--- Benchmark complete
On Server
reps 20 rounds 1000
UnCached 2.0708000000838e-07
Cached 2.1247999993221e-07
The Code: files/getconvar_cache.lua
This was tested with the inbuilt math.pow function and also the default power-of operator.
TL;DR: Using the "roof" operator is slightly faster in gmod. The speed is the same in Lua5.4.
The result (average of 100 rounds with 10000 calculations each):
GMOD SERVERSIDE CONSOLE
x^2 math.pow: 4.0210000111074e-06
x*x: 3.7120000251889e-06
x^2 ^: 3.5500000080901e-06
LUA5.4 CONSOLE
x^2 math.pow: 0.00015625
x^2 ^: 0.00015625
x*x: 0.00015625
The Code: files/x_squared_vs_x_times.lua
TL;DR: A for loop is way faster than while.
The result (100x 10000):
GMOD SERVERSIDE CONSOLE
while: 0.00000957507999874 (9.57507999874e-06)
for: 0.0000024306500009516 (2.4306500009516e-06)
LUA5.3 CONSOLE
while: 0.00015975869999999
for: 0.000092178400000006 (9.2178400000006e-05)
The Code: files/while_vs_for.lua
TL;DR: Defining the color once is faster than using Color() in a loop.
The result (1000x in SysTime() time taken):
Local: 0.0000070583999997496 (7.0583999997496e-06)
Non-Local: 0.0000091475999998067 (9.1475999998067e-06)
The Code: files/local_color_vs_redefining.lua
TL;DR: Creating a lot of global variables neither slows down the game nor creates any noticable impact on performance.
I heard from multiple sources that "creating many global variables is bad".
I tested this by creating 10.000 global variables and observing the game afterwards. An example with the runtime of 10.000 math function calls:
Time _G-count time taken
Before 3133 0.00044090000005781
After 13144 0.00043160000018361
The FProfiler statistics and general gameplay feel didn't change either.
For more details on the test (or if you want to verify it yourself) check out my script I used:
files/global_variables.lua
TL;DR: It is better to always use the # operator instead of table.Count for sequential tables.
This comparison only works on sequential (or numerical) tables. These are tables that have a numerical, rising number as a key, in other languages you would call this "list" or array.
Table.count would always be used if you want to count a string-index table (also called dictionary or map), but in cases like this both can be used.
Result:
# Table size: 1000
Hashtag: 6.2949998141448e-08
TbCount: 1.5430600009495e-06
# Table size: 10
Hashtag: 4.4980000029682e-08
TbCount: 9.0410000939301e-08
As you can see above, the table.Count method is getting slower the larger the table becomes. This is because it loops through all the elements in the table and counts them, according to an old edit from the gmod wiki.
The Code: files/table_count_vs_hashtag.lua
TL;DR: It is faster to randomly generate an index of a sequential table and get a random value this way instead of using table.Random().
Result:
--- Benchmark complete
On Server
reps 20 rounds 10000
small old 4.6385350004698e-07
small new 1.2625500000013e-07
big old 3.1202909999735e-06
big new 1.2906200004807e-07
The Code: files/table_random_overwrite.lua
TL;DR: It is not faster to overwrite the table.Count function and count the table yourself in LUA.
Result:
--- Benchmark complete
On Server
reps 20 rounds 10000
count small old 2.7576250000664e-07
count small new 2.6563899996773e-07
count big old 1.8296299999935e-06
count big new 1.8260360000758e-06
The Code: files/table_count_overwrite.lua
TL;DR: It is not faster to overwrite the file.Write function to use lower level lua functions to write files.
Result:
--- Benchmark complete
On Server
reps 20 rounds 100
write old 0.00076962965000047
write new 0.00079238309999975
The Code: files/file_write_overwrite.lua
TL;DR: It is faster to get variables from a table than from the ply object. Setting a variable has nearly the same speed though, only getting the variable is slower on ply.
I and many others use code like ply.lastSpawned = CurTime(). This is quite easy to use and understand, but in performance aspects it is slower than using a table. An example with a table would be tab[ply]["lastSpawned"] = CurTime().
This is because reading this variable from a player uses a (slower) __index method. If you ever used FProfiler then you would probably have seen this "metamethod __index player" function near the top of "most runtime in ms".
I tested this by setting and getting random numbers with both methods 10.000 times, and I ran this 1000 times and calculated the average time taken. Result:
ply: 0.0033284901999953
tab: 0.0016247142000013
As you can see, it is quite faster (by 2x) with a table.
Using a table also has its downsides, like having to clean it up regularly to avoid ram leaks and having to check 2 table indexes each time you want to verify it.
Source code: files/ply_vs_table_index.lua
For the full documentation please look at https://shira.at/gmod/mysql_sqlite.html.
Setup: MariaDB on Debian 10 installed on the same server as the Garry's Mod Server. Tested with DarkRP's MySQL functions and the mysqloo module.
The time is measured from first requesting the data until we can use the data.
Summary: MySQL is slower in comparison to SQLite. It needs an external connection and needs to wait for request/response delays either via network or via internal linux sockets. It's great for synchronisation of multiple servers, but it's not a "must-have" for servers.
To get the delay for a query without read or write operation we can query SELECT 1+1.
Result:
--First Test
SQLite ResponseTime: 0.000045883003622293s
DRPMySQL ResponseTime: 0.020215623022523s
--Second Test
SQLite ResponseTime: 0.000027909001801163s
DRPMySQL ResponseTime: 0.049109992978629s
Query: SELECT rpname FROM darkrp_player WHERE uid = 76561198071737444
Result:
SQLite Read Time: 0.00016302999938489s
DRPMySQL Read Time: 0.048014674001024s
Query: INSERT INTO darkrp_player VALUES(123123,'Test User',1,2)
Result:
SQLite Write Time: 0.00020395400133566s
DRPMySQL Write Time: 0.06270497100013s
DarkRP's SQLite method uses sql.Begin and sql.Commit for SQLite based bulk imports. This makes it way faster than doing 1000 inserts after each other.
DarkRP's MySQL method simply puts the queries into a table and sents them one after the other to the mysql-server. This is way slower, as the time neede is simply the time one insert takes times 1000.
Result:
SQlite 1000 Inserts: 0.004849937000472s
DRPMySQL 1000 Inserts: ~30.000000000000000s (calculated)