The objective of this investigation is the possibility of building a model of physically based rendering in a cloud computing system. Analyzing various algorithms for rendering process, and various ways to construct a distributed rendering system. There is an investigation of reaching cloud computing properties as high availability, fault tolerance, auto scaling, and exibility in rendering domain. This paper describes various solutions of the problems of construction with the use of rendering algorithms in distribution context. The paper opens up possible ways to build a distributed rendering system and directions for future research.
The cloud compute technology is a method of reaching high quality rendering
system. The purpose is not related to the theory of quality and other
formalization of computing software characteristics. High quality is a useful for end user
process these days. The physically plausible 3D graphics has been in demand for
the last years. Various elds use photo-realistic graphics for the solutions. For
example, rendering of virtual furniture for demonstrations goals [
The city of San Fransokyo is illuminated by more than 216,000 street lights and it has 83,000 procedural-built buildings with an equal number of street props and trees. On this paper, some algorithms are considered from the render system for a big scene from such cartoon, Big Hero 6. Such system was developed by R&D, a department of Disney company, that shows many methods for outcore rendering. The system is monolith in-house solution and doesnt have cloud properties by design. Such characteristics as exibility, reliability, availability are needed for e ciency of rendering a big scene for science and production tasks. It enables to construct end-user abstraction as SaaS or PaaS. The big problem is on how to construct computational model with such property for physicalbased rendering pipeline. The paper contains review of existing experience and suggested novel methods, and directions for future researchers. 2
There are three main rendering pipelines. A rasterized rendering pipeline, a ray tracing pipeline, and a hybrid pipeline.
The rst pipeline de nes rasterization as an image synthesis system. The place
of this pipeline is between the processing of primitives and the processing of
fragments (in Figure 2, the rendering pipeline is based on rasterization).
Rasterization is the projection and processing of simple geometric primitives, mainly
given in the form of a triangular grid on a fragment of a raster screen. A fragment
is a part of the image obtained as a result of rasterization of the primitive, and
might not be an element of the light calculation. The method of rasterization
is most e ective for constructing simple images, but its a disadvantage in the
impossibility of correctly calculating the indirect illumination. Rasterization is
not a physically based solution and is usually used in the production of simple
video games. It is usually not used to create cartoons and special e ects in the
lm industry, and, especially for physically correct calculations. Therefore, the
rendering pipeline, based on the principles of rasterization, does not meet the
requirements.
It is known that the architecture of a distributed rendering system places high
demands on the e ciency and correctness of modeling. Therefore, the most
universal methods of forming photorealistic images are the methods based on ray
tracing. The methods have their advantages and disadvantages. The simplest
variation of the ray tracing method is based on the approach of visualizing
sources of primary and secondary scene brightness, i.e. ray tracing (or
visualization paths) from the observer's eye to the geometric objects of the scene (Fig.
3. shows the main algorithm for backward ray tracing) [
Another method - the path tracing method - is the generalized ray tracing, which is closest to the physical laws. Rays are traced until they are completely absorbed or leave the stage. In addition, for each scattering event, the direct brightness is calculated, which allows to signi cantly increase the overall rendering e ciency. However, the e ectiveness of this method is not always optimal in scenes with complex, for example, caustic lighting and in the presence of a large number of primary light sources.
The bidirectional ray tracing method combines two methods: path tracing
and Forward Monte Carlo Ray Tracing methods. In the rst case, the trace
comes from the eye of the observer, and in the second case from the light source.
By connecting the two paths with shadow rays on the di use surfaces of the
scenes, it is possible to estimate the brightness of the secondary illumination
for the screen point for which the ray was emitted from the observer. The main
problem of this method is the calculation of caustic lighting and the
corresponding caustic brightness, since this method is based on the presence of at least two
di use events on the paths of the direct and reverse rays. The main advantage of
this method is its unbiased in assessing the brightness of images of screen dots.
The method of photon maps is a fairly simple and e ective method of
photorealistic visualization. Its main disadvantage is a possible bias in the estimate of
the average brightness of the image point. The main idea of this method is to
build an irradiance map (photon map) by tracing photons from a light source and
using the resulting map to calculate the brightness of the indirect and caustic
lighting. When the ray crosses the photon map element, the brightness is
calculated, which accumulates at the corresponding points of the image [
The hybrid pipeline combines two methods: rasterization and ray tracing. This allows you to improve graphics in real time and to add additional visual e ects obtained by tracing a small number of rays. An example of this approach is to use rasterization to calculate direct illumination, that is, for the rst generation of the rays. The small amount of traced rays a ects the quality of the nal image and cannot provide the physically correct modeling needed for lmmaking when 3D-scene images are built for weeks (for a small movie plot) on supercomputers, as what Disney has done. In conclusion, we can say that at the moment only the pipeline, built on the basis of ray tracing, allows us to solve the problems under consideration. 3
On this paper, classic distributed network pipeline means render farm. Such clean concept of network render architecture doesnt have such cloud property as SaaS or PaaS system for reliable and available rendering of big scene. The main questions here are related to fault tolerance, scaling, and con gurable pipeline for rendering big scene. System needs to work after failure more than half of system resources and can full recovering process. Its important for rendering of big scene when computation time takes weeks. Scaling is about dynamic reconguration pipeline. To make it faster, simple scaling of resources is the solution and this wont restart the whole pipeline. on gurable pipeline is important for economic pro t. It about the ratio of cost to time and quality. The di erent task as cartoon or simulation for scienti c research have di erent requirements and di erent cost with time. The next part reviews of ways to go out from distributed network pipeline limitations and solve considered problems. The rst step for solution is decomposition of the system where the whole pipeline is on one host. The unit of decomposition is micro-service. Its a small program module that is independent and autonomous. Such module can replace any other module of corresponding class. Also, it supports horizontal scaling through increase count of micro-service. There are three main class of modules: pre-processing microservice, ray tracing processing micro-service and post-processing microservice. Pre-processing is consisted of decomposition of scene on voxels and initialization of pipeline. Decomposition allows to distribute pipeline on micro-service, when one service per voxel. Algorithms of voxelization are many. One of corresponding methods is binary space partitioning. This is a method for recursively subdividing a space into voxels. Its process are continuing while voxels has big size for render on one micro-service. As a result, there is hierarchy of objects for ray tracing processing. Each voxel has type as empty voxel for part of scene without object. It optimizes processing. Ray tracing processing includes one of methods for lighting processing like photon mapping or metropolis light transport. That service has uni ed interface and there is no meaning concrete method for tracing. Each service can send to other services batch of rays for processing. The process is called streaming. It is the key process for rendering pipeline. It enables the process of out-of-core scene also. The problem of fault tolerance is solved by replication of produced data. Produced data is stored to database server for reliable. When a database server is breaking, one of replica become to main and continuing processing (Fig 4. shows big picture of system).
That architecture pattern is called service mesh. Post-processing are forming image, estimate accuracy, construct predictions, doing ltering, and compressing for output artifacts. If that is cartoon then combining. The main issue here is the process of big data output to send an image and cartoon and video to user. There is distributed ltering and processing for solving. As considered important process is streaming between ray tracing services. In the next section reviews reliable solution.
The worst case of rendering is when a scene needs more than one host. To do this, there are many methods as out-of-core geometry and out-of-core ray tracing or out-of-core texturing. In engineering, the method is usually making a stream of data when the data is too big. To reach cloud property, it is necessary to make exibility, reliability, availability and scaling stream. The problem is solved in a system of big data streaming called Apache Kafka. That system that is a queue solved problems of reliable distributed streaming. It allows transmit a big data through network in reliable and scaling way. Usually Kafka used in pipeline for big data processing (Fig 5. shows common big data pipeline). That experience actual for distributed rendering system starting from persistent layer to streaming and processing. For persistence layer, there are variants such as Apache Ignite, PostgreSQL and Apache Cassandra. The estimation of e ciency on these systems in distributed rendering pipeline, is a separate subject for future researches.
Constructed computational model with service mesh pattern enables to estimate various algorithms of various parts of the pipeline. Distributed architecture of system enables to replace any algorithm thus the system will evaluate on time. 5
A study was conducted to investigate other possible ways of organizing cloud computing with distributed processing of the renderer, and the problem of separating a large process between components of a cloud distributed system was also considered. The study showed that the cloud rendering system was being developed will allow solving complex computational problems related to the assessment of visual discomfort caused by the mismatch between the vergence and the accommodation of human vision. With the help of the developed highperformance computing system, it will be possible to perform time-consuming and resource-intensive calculations necessary for synthesizing images generated by virtual prototypes of the human vision system and virtual and mixed reality systems, and for assessing the consistency of vergence and accommodation of human vision in the resulting images. This work was partially funded by the RFBR grant No. 18-08-01484.