1. Introduction
1.1. Preface
As the network technology evolved, MMOG (Massive Multiplayer Online Game) has become the one of the largest computer game industry. Each MMOG forms a virtual world, within which people act and live together. No matter where they are, the virtual world connects them in the same space, at the same time.
1.2. Motivation
MMOG is sweet and sour for most game company. The sweetness comes from the fact that no private copy is possible because everyone needs to connect to the virtual world, thus increase the revenue. On the other hand, the sourness is high cost and long development cycle. Because there are several critical aspects needed to address even for a simplest MMOG, such as networking which requires enterprise-level reliability, stability and scalability to supply tens of thousands player concurrently.
To ease the overall MMOG development, several middleware solutions exist both in commercial market and open source community. They all have the same goal to support the online game development and to shorten the development cycle by reducing the building complexity. Although there is no in-depth comparison between these solutions, but generally speaking, in terms of scalability and robustness, they are not good enough as what we expect. According to some statistics, all real-world online games fail to scale up to more than 200k players, so that the game operator
2
usually gives more shards (i.e. a separate virtual world) as the number of subscriber grows. Each shard is running on a server cluster consists of about 5 to 20 servers to form a virtual world with thousands of concurrent users. The number of concurrent players on each server is about hundreds to thousands, which is pretty small.
Meanwhile, players in different shards cannot communicate or interact with each other because they are physically separated world and this makes shard design less-attractive.
As shard is less-attractive, there are some research works and middleware trying to break the shard boundary to make it a shard-less design. For example, Sun Microsystems initiate an ambitious project called Darkstar [1] to build a large and shard-less virtual world, but the project doesn‟t seems to actually solve the problem from their latest benchmark numbers, and also the cost for entire server cluster still remains high and unaffordable for small game companies.
In this thesis, we look into the scalability problem existed in most commercial and open source middleware solutions, and propose a solution to scale up to a very large number of concurrent players in single server while maintaining the ease of use in overall game development.
1.3. Problem Description
Designing a flexible, scalable, customizable, and also easy-to-use MMOG platform is challenging. Among all, the most critical factor of a MMOG platform is still the performance which refers to scalability. Scalability usually a key of cost for operations because the better scalability a platform has, the fewer clusters operator needs to deploy. In additional to scalability, flexibility and ease of use of a MMOG platform are also significant. Flexible and customizable designs would give more
3
possibility for game designer to construct more complicated game play, which can really make differences in competitive online game market.
To give better understanding of the problems, we first elaborate some design pattern and common constraints in MMOGs. Almost all MMOGs are based on command-update communication model in which every players uses a client application to send commands to the server and receive updates in return. For each command, server needs to process the request through a specific game logic and sends updates back in a limited timeframe. The timeframe depends on game genre, and, for example, most MMORPG requires 200ms response time to deliver a smooth game play, and MMOFPS usually requires less 70ms to make it playable. Many research works have been conducted to reduce the latency, but most of them attack the problem from the network perspective. They tried to reduce the network latency by proposing different network topologies. Proposed communication architecture includes peer-to-peer and scalable server/proxy architecture. They have succeeded to alleviate the network latency between server and client. However the number of concurrent player per server is far from 10,000 even if the network latency has been improved by various works.
In our point of view, as the network technology evolved, the network latency is getting lower and lower and has dropped to a certain level. The bottleneck becomes the client command processing on server-side. As the number of players grows, the number of client commands needed to be processed in the limited timeframe increases.
However, the current design of CPU is not capable of processing such a huge amount of commands in the limited timeframe. Although recently multi-core CPUs are introduced, the cache-misses occurred for different execution context of game logics and network handlers, and the synchronization among threads or processes will even
4
lower the CPU processing capability of client commands when lock is needed to avoid update conflict to guarantee consistent results.
Because of the limit, we started to investigate the source the problem and found it the result of sequential nature of CPU processing. In our previous work, we give the answer for this problem: make it parallel by using GPU. For the last decade, GPU had evolved into a powerful array of general SIMD processer instead of just being a simple 3D accelerating silicon. To date, the computation power of GPU is hundreds of times more than that of CPU, and it‟s believed that the GPU computation power will continue to supersede Moore‟s Law, while CPU has hit a barrier.
Although light has been spotted on GPU, converting a sequential program to a parallel one is not straightforward. Algorithm needs to be re-designed and data structure needs to be re-organized to reflect the different access pattern in parallel programming. Several attempts have been made to exploit GPU computation power to general problem such as collision detection [2] and online database processing [3].
For MMOG, the world first GPU-assisted MMOG platform is proposed in [14].
However it is still not a “full-featured” platform in terms of flexibility and ease-of-use.
Due to the fact that most programmers or game developers are not familiar with parallel programming or GPU programming, it‟s still not possible to build an online game server that is accelerated by GPU computation.
1.4. Research Objectives
In this thesis, we try to give a more modular GPU-assisted MMOG platform, and provide way for game designer to construct game logics processed on GPU without the knowledge of parallel programming or GPU programming. As we don‟t require game designer to write real GPU code, an interface that bridges game logic design
5
concept to real code is needed, which is basically a code generator. In addition, various optimization of code generator needs to consider further enhance the server performance , thus increase the scalability of the platform.
1.5. Research Contribution
We discusses several issues and problems to build a MMOG platform with flexibility, scalability and customizability while handling client commands and create updates on both CPU and GPU. Because of the architecture difference between CPU and GPU, platform itself has been re-designed to exploit the CPU/GPU computation power by generating compatible code on both CPU and GPU. We define a set Extensible Markup Language (XML) schema for game designer to customize their game logic and provide a code generator to translate the XML to GPU code pieces as the server plug-ins to process the entire game play. A number of regression tests are performed to validate the correctness and robustness of the code generator and demonstrate the efficiency of GPU-assisted MMOG platform. As a result, commands issued by MMOG clients can be processed efficiently by GPU-assisted server through the generated GPU code in parallel.
6