How to optimize performance on Roblox — Transcript

Full transcript

1. 0:00Hey everyone, it's Andrew from Roblox.
2. 0:02This video looks at common performance
3. 0:04problems in Roblox experiences and
4. 0:06strategies you can use to mitigate them.
5. 0:08Before we dive in though, two quick
6. 0:10disclaimers. One, I am not an
7. 0:12environmental artist. At several points
8. 0:14in this video, I'll ask you to use your
9. 0:15imagination because I don't have the
10. 0:17time or the skill to build the things
11. 0:18I'm talking about. Two, I'm not
12. 0:21advocating for compromising on your
13. 0:23artistic vision. Build the experience
14. 0:25you want to build. But along the way,
15. 0:26you might consider how these strategies
16. 0:28can make your experience more performant
17. 0:30and help it reach a broader audience.
18. 0:32With that out of the way, let's start
19. 0:33with rocks. That's right, rocks. Lots of
20. 0:37rocks. We're using rocks to demonstrate
21. 0:39a fundamental concept of computer
22. 0:41graphics. Draw calls. Draw calls are
23. 0:43commands from the CPU to the GPU to
24. 0:45render something on screen. If you have
25. 0:47lots of unique meshes, the Roblox engine
26. 0:49has to make draw calls for each mesh. On
27. 0:52mobile, you can quickly get into a
28. 0:53situation where your device is
29. 0:54struggling to keep up. If you were to
30. 0:56look closely at the beautiful stone
31. 0:58walls of Ireland, you'd of course see
32. 1:00lots and lots of unique rocks. But what
33. 1:03if you constructed a wall by duplicating
34. 1:04four or five rocks over and over and
35. 1:06then artfully arranging, coloring,
36. 1:08resizing, and rotating them? Unless
37. 1:10somebody really scrutinized that wall,
38. 1:12would they even notice?
39. 1:15A technique known as batching lets your
40. 1:17CPU make a single draw call for tons of
41. 1:19objects as long as they share the same
42. 1:21mesh. So you can import a small handful
43. 1:23of particularly excellent rocks and use
44. 1:25them to build all sorts of things.
45. 1:26Walls, landscapes, cliffs, rivereds,
46. 1:29whatever. Here I've imported the same
47. 1:30rock many, many times so that each one
48. 1:32gets its own mesh ID. Don't do this. I
49. 1:36only did it to illustrate the point.
50. 1:38Look at all those draw calls. If I
51. 1:40delete all of those duplicate imports
52. 1:41and just copy the same rock over and
53. 1:43over, draw calls don't increase at all
54. 1:44because I'm reusing the same mesh. The
55. 1:47goal here isn't to keep your draw calls
56. 1:49needlessly low and build bland
57. 1:50environments. Instead, this sort of
58. 1:52optimization gives you the headroom to
59. 1:54use your draw calls on the things that
60. 1:55really matter. On a low-end mobile
61. 1:57device like this one, it can be the
62. 1:58difference between a game that looks
63. 1:59like this, running pretty darn well, or
64. 2:02a base plate with some rocks struggling
65. 2:04to stay above 30 FPS. But you don't have
66. 2:06to take my word for it because I've
67. 2:07recruited the one and only Mr. Chicken
68. 2:09Rocket for a quick discussion of draw
69. 2:11calls and relative device performance.
70. 2:13Take it away, Peter. Hey, thanks Andrew.
71. 2:15When I create a game, I usually start
72. 2:17with a basic budget. Something like half
73. 2:19a million triangles and 500 draw calls
74. 2:21in the scene for any given view. And
75. 2:23then I stick to that budget as I build.
76. 2:24It's the best way I know of to avoid
77. 2:26having to go back and redo work because
78. 2:28things aren't running at 60 frames per
79. 2:29second anymore. And I hate redoing work.
80. 2:32Staying under budget while I'm working
81. 2:33in studio means I can be fairly sure
82. 2:35that it'll run great on a wide range of
83. 2:37devices, but every now and then I'll
84. 2:39still break out an old iPhone and go
85. 2:41check. When I am testing on device, what
86. 2:43I'm usually looking for is making sure
87. 2:44that it's still running at 60 frames a
88. 2:46second. And sometimes I'll check the
89. 2:48micro profile to see what's going on
90. 2:50with CPU usage, too. Next, let's talk
91. 2:53streaming. Streaming is enormously
92. 2:55complex under the hood, but it's easy to
93. 2:57enable and offers huge performance
94. 2:58benefits. In essence, streaming creates
95. 3:01an invisible cube around each character.
96. 3:03As characters move around, parts stream
97. 3:05in and out based on whether they're
98. 3:07inside or outside of the cube. Streaming
99. 3:09improves frame rate and join times.
100. 3:11Maybe most importantly, it reduces
101. 3:13memory usage on the client, which can
102. 3:15help you build more expansive worlds,
103. 3:17minimize crashes on low-end devices, and
104. 3:19let more people play your game. The
105. 3:21default values are good options for a
106. 3:23wide variety of experiences, but
107. 3:24streaming has a bunch of properties you
108. 3:26can tune. Check the documentation for
109. 3:27full descriptions of these properties.
110. 3:29The key ones for today are streaming
111. 3:31target radius and streaming min radius.
112. 3:33Streaming target radius defines the
113. 3:35outer bounds of the cube, the maximum
114. 3:37distance at which players see the full
115. 3:38detail of your experience. If your
116. 3:40experience relies on longer viewing
117. 3:42distances for visual impact, consider
118. 3:44bumping this up a bit as you tweak the
119. 3:46value. Test your experience and note the
120. 3:48impact on memory usage and frame rate.
121. 3:50There's almost always a sweet spot that
122. 3:52offers tangible performance gains
123. 3:53without diminishing visuals. Streaming
124. 3:56min radius defines a second smaller cube
125. 3:58inside the larger one. Parts inside of
126. 4:00it have the highest priority when it
127. 4:02comes to streaming. Because the engine
128. 4:04is already trying to stream in content
129. 4:05at the streaming target radius, this
130. 4:07second cube is more of a fallback for
131. 4:09situations in which the engine is
132. 4:10struggling to keep up and has to
133. 4:12prioritize parts closer to the
134. 4:14character. As such, be sure to keep this
135. 4:16value smaller than streaming target
136. 4:17radius. If the two values are the same,
137. 4:19what you've done is make everything a
138. 4:21priority. And when everything's a
139. 4:22priority, nothing's a priority. Yet
140. 4:25again, you don't have to take my word
141. 4:26for it because I've recruited
142. 4:27Wellblander from our engine team to say
143. 4:29more. Hey, thanks. As mentioned, I'm
144. 4:31Well, Blander, a PM on the streaming
145. 4:32team here at Roblox, and I want to
146. 4:34quickly talk about how streaming can
147. 4:36help with humanoids in your games. In
148. 4:39short, humanoids can really stress your
149. 4:41players devices, especially those
150. 4:43lower-end phones and tablets some might
151. 4:44be using. Humanoids are animation and
152. 4:47physics heavy, causing high CPU usage
153. 4:50while also using lots of memory and
154. 4:52network bandwidth. And in really large
155. 4:54worlds with lots of players and/or NPCs,
156. 4:57your players devices, no matter their
157. 4:59strength, will really struggle to keep
158. 5:01up with all your game's characters. Hey,
159. 5:04well, streaming can help here, too. If
160. 5:06those humanoids are in the distance and
161. 5:08they aren't impacting gameplay, then
162. 5:11chances are, well, they're not needed.
163. 5:13With streaming on, the Roblox engine
164. 5:15will automatically stream these distant
165. 5:18characters out, saving your players
166. 5:20device resources, resulting in smoother
167. 5:23gameplay and fewer
168. 5:25crashes. But well, what if there's a
169. 5:27particular humanoid that is crucial to
170. 5:29your gameplay? I don't know. Um, take
171. 5:32for example some opponent player
172. 5:34carrying a captured flag or something,
173. 5:36but they're really far away. Well, then
174. 5:39you can use StreamY's persistence option
175. 5:41to ensure that flag carrying character
176. 5:44is always streamed in. Pretty cool. But
177. 5:47beware, if you overuse persistence,
178. 5:49you'll find yourself right back dealing
179. 5:51with the problems that streaming was
180. 5:53meant to solve in the first place. So,
181. 5:55use this option sparingly. Last, I want
182. 5:58to say that sometimes even streaming
183. 6:00isn't enough and instead you might need
184. 6:02to design your places and gameplay to
185. 6:05space players out or reduce the maximum
186. 6:07numbers of players per server. If you
187. 6:09think your experience might have issues
188. 6:11here, keep an eye on the update
189. 6:13invalidated fast clusters tag and the
190. 6:16micro profiler. Thanks so much, Wland
191. 6:18Blander. Finally, let's talk central
192. 6:20processing units, CPUs. If you owned any
193. 6:23of the original smartphones, you might
194. 6:24have fond memories of how revolutionary
195. 6:26they felt. But you also might remember
196. 6:28that they were so slow. Smartphones have
197. 6:30made extraordinary strides over the past
198. 6:3215 years. But due to thermal
199. 6:34constraints, their CPUs are still only
200. 6:36about one quarter the speed of powerful
201. 6:38desktops in longunning intensive
202. 6:40workloads like, you know, games. All of
203. 6:42this is to say that it's easy to create
204. 6:44experiences that run fine on your
205. 6:46development machine and completely tank
206. 6:48CPU performance on a smartphone. There
207. 6:50are three main culprits here. Scripts,
208. 6:52physics, and lighting. Let's start with
209. 6:55scripts. Placing code into run
210. 6:57service.preder or its deprecated
211. 6:59counterpart render step causes it to
212. 7:01execute before each and every frame.
213. 7:03With a budget of only 16.67 milliseconds
214. 7:05per frame in a 60fps experience, even
215. 7:08seemingly minor additions to pre-render
216. 7:10can have an enormous impact on
217. 7:12performance. See here how I was able to
218. 7:14bring this experience to a crawl with
219. 7:16just a few lines of code by searching
220. 7:17the scene every frame and doing a bit of
221. 7:19ray casting. This isn't to say that you
222. 7:21should never add code to run service.p
223. 7:22pre-render. Sometimes it's the best
224. 7:24solution, but you should keep it to the
225. 7:26bare minimum, such as custom camera
226. 7:27updates or other elements that require
227. 7:29precise visual synchronization. Whenever
228. 7:32possible, write code that runs in
229. 7:34response to specific events rather than
230. 7:36before every frame. If you want a call
231. 7:38back to run on a loop, consider whether
232. 7:40it really has to happen every frame or
233. 7:42whether it can happen more infrequently.
234. 7:44Next, physics. If you see the physics
235. 7:46stepped tag in the microprofiler
236. 7:48becoming a limiting factor to your frame
237. 7:50rate, check which parts in the scene
238. 7:52really need to be unanchored. The sheer
239. 7:54number of simulated parts is the most
240. 7:56common cause of poor physics
241. 7:57performance. You should also consider
242. 7:59enabling adaptive physics stepping in
243. 8:01the workspace, which lets Roblox run
244. 8:03certain physics calculations less often.
245. 8:06Finally, lighting. Lighting and shadows
246. 8:08are critical to the atmosphere of an
247. 8:09experience. The client eventually falls
248. 8:11back on simplified voxil technology. As
249. 8:14graphics quality level decreases, so
250. 8:16keeping lighting performance high means
251. 8:17that more players get to see the scene
252. 8:19as you designed it. Consider setting
253. 8:21cast shadow to false on small parts
254. 8:23where the shadow would barely be visible
255. 8:25anyway. Also, watch out for movement. If
256. 8:27you have a large light that's always in
257. 8:29motion, the engine has to continuously
258. 8:31calculate the effect of that light on
259. 8:32its surroundings. See here how CPU
260. 8:34performance dips as the lights approach
261. 8:37and then improves as they depart. I know
262. 8:39what you're thinking, Andrew, no. You
263. 8:41didn't find a third performance expert
264. 8:42wandering the halls of Roblox, did you?
265. 8:44Well, I did. Go for it, Owen. Hi all.
266. 8:47One thing Andrew didn't mention is you
267. 8:49can actually create your own tags when
268. 8:50profiling your experience. Using the
269. 8:52debug.profile begin and debug.profile
270. 8:55end methods, you can tag a section in
271. 8:56your code with an arbitrary string. Then
272. 8:58when you pull up the microprofiler, you
273. 8:59can look for these tags to reason about
274. 9:01the impact of a specific game system.
275. 9:03Please remember though, code is supposed
276. 9:04to take time to execute. If you're
277. 9:06encountering frame rate hitches or your
278. 9:08players are complaining about these
279. 9:09issues, perhaps low heartbeat, you
280. 9:11should definitely dive in and diagnose.
281. 9:12But early on in development, you really
282. 9:14don't want to spend too long obsessing
283. 9:15with how many milliseconds each function
284. 9:17in your code takes to run. All that time
285. 9:19you'll spend micro optimizing is time
286. 9:21you're not spending making your game
287. 9:22more fun. And over time, all these micro
288. 9:24optimizations you add to your code will
289. 9:26add up. You have to stay focused on the
290. 9:27big picture. Performance optimization is
291. 9:30an all-encompassing, extraordinarily
292. 9:32challenging topic, so there's no way we
293. 9:33could cover every possible facet in a
294. 9:35video. I want to extend a huge thank you
295. 9:37to our three experts for joining me
296. 9:39today and shedding some additional light
297. 9:40on the subject. If you want to dive
298. 9:42deeper into performance, including way,
299. 9:44way more information on tools, memory
300. 9:46usage, streaming, humanoids, rendering,
301. 9:47and everything else from this video,
302. 9:49check out the documentation at
303. 9:52create.roll.com. Thanks for watching.