2007 Microsoft Corporation. All rights reserved. Microsoft, Windows, Windows Vista and other product names are or may be registered trademarks and/or

Transcription

1 2007 Microsoft Corporation. All rights reserved. Microsoft, Windows, Windows Vista and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries. The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION. 1

2 This talk will address: The five components that make up EnSight Client / Server design SOS for big data RLSOS for even bigger data DR for big rendering or VR Choosing where to run the EnSight components How to simplify launching everything in modern computational environments Miscellaneous tips for working with big data Describe my responsibilities at CEI and briefly my background. 2

3 Most users of EnSight simply see this view of EnSight: the graphical interface running on their desktop. Even though EnSight may be running entirely on the user s desktop, there is always an EnSight Client application and an EnSight Server application. The Client and Server work together to provide all of the functionality that is EnSight. 3

4 What do the Client and Server do? Why should I care? (next 2 slides) 4

5 5

6 6

7 Back in 1987 there wasn t really a way to visualize supercomputing data. Looking at rows of numbers on a terminal (or Mac) was about the only good method along with XY plots. John Kogut at NCSA with a Mac and a Cray XMP24 (1987). In 199x Cray introduced the software package MPGS as a way to look at 3D visualizations being produced on Cray supercomputers. The computation was done on the supercomputer and the visualization was done on a Silicon Graphics workstation. MPGS was the the predecessor to EnSight. The graphics people at Cray left to start CEI Inc. in EnSight (and formerly MPGS) has always been a Client / Server application utilizing the strengths of remote computation resources and high performance graphics workstations. 7

8 Why should I care about big data? If analyzing your problem is very fast and easily with EnSight, then maybe you do not. However, most scientific and engineering companies are usually wanting to grow their simulation size. Additionally, due to globalization, a company may choose to run their simulations around the world thus requiring users to utilize remote computational resources. Because of this, running EnSight distributed may be required. Some of the same issues for big data apply to distributed computing. 8

9 Big is always a shifting target. It depends on current computer configurations vs. simulation size. Basically, too big is when the simulation is too big for the user s desktop/laptop. Desktop and laptop computers keep getting more powerful. However, some companies take the opposite strategy by shrinking desktop capability to reduce IT costs. This is unfortunate for engineers since they really could benefit from relatively inexpensive workstations with lots of RAM and modern graphics cards. These can be more powerful than supercomputers in the recent past! Nonetheless, EnSight can run well in both environments (and others). This is one of EnSight s strengths. It can easily adapt to the computational environment that customers use. When the user s desktop/laptop is not big enough, then move the EnSight Server to another computer. Typically, this is also the computer that ran the simulation program but not always. 9

10 What if one EnSight Server is not enough? Use more! This is EnSight SOS. The EnSight SOS process sends command from the Client to all Servers. Then it gathers results from each Server and sends them back to the Client. Add some rough metrics. Need to benchmark YOUR computer systems to find out what works well on YOUR computers. There is not one answer. 10

11 Okay. You want to use multiple EnSight Servers. Where should you run the multiple Servers? It might be sufficient to run them all on one SMP if it has sufficient numbers of CPUs and RAM. Alternative, you might run them on a cluster. Any combination is possible. It helps to understand basic performance optimization and scalability, though, if you wish to optimize for very challenging problem sizes. We can help. The SOS should run where it has the best network connectivity between and and the Servers. 11

12 CEI frequently gets asked about using threads to improve performance. Computer companies like to sell computers with many CPU cores in one box. From a high level threading seems like a very easy answer to scalability. Unfortunately, this is not so simple. Thread scalability is problematic past 4 to 8 threads (for any program unless specially rewritten). Threading requires careful management of memory access, cache access, and CPU assignment and thread migration. It can be very difficult to scale beyond a few threads. It typically requires that a program be carefully rewritten with data access carefully managed. Furthermore, threading does not help at for running on a distributed memory cluster. To use a program on a cluster it must be explicitly written to run on different computers. Fortunately, programs that work well on clusters also tend to work well on SMPs. Thus, write the program well for clusters and don t worry too much about threading. This is the approach of EnSight. Since the SOS model explicitly partitions memory and CPU access, it scales well on modern SMPs *AND* clusters. EnSight does use threads in limited situations where it makes sense such as EnSight calculator operations and other basic numerical operations used throughout EnSight. This was greatly improved in EnSight 10.1 and newer versions will add more thread support. BUT! SOS will always give better general scalability beyond a few cores. EnSight can use threading and SOS at the same time. This is encouraged. Additionally, large SMPs (many cores in a single box) tend not to perform as well due to mismatch CPU performance vs. cache/memory/disk IO limitations of a bus-based architecture. Much better to have a cluster of smaller SMPs each with lots of RAM and large caches. See the EnSight How To manual for the environment variables that affect EnSight threads. 12

13 This is new to EnSight It will be fully tested in EnSight EnSight RLSOS may also give better performance when using more than 64 Servers since it fattens the communication network allowing greater concurrency in computation and communication. If not using 64 or more Servers, then do not consider RLSOS. CEI recently did a test on a transient data set containing 8 billion elements using 200 Servers and RLSOS. It worked perfectly and quickly. Not many years ago this would have been considered a Hero Problem. Now it is easy. 13

14 For a *very* rough estimate, 1 DRClient for every 10M-30M visible polygons using current OpenGL cards. A single high-end OpenGL card will ALWAYS perform better if it handle the number of visible polygons. Network latency and bandwidth greatly affect real-time rendering rates which is why parallel compositing (EnSight HPC+) is best if there is no other option. But, if there is another option such as a faster graphics card, then that will probably be better. 14

15 The same technology (and same EnSight components) used by EnSight HPC for parallel rendering are also used by EnSight VR. They are separate products, though, and have different options for configuring use for HPC vs. VR. EnSight VR is for virtual reality systems: Caves, Raves, PowerWalls, etc. It also supports 3D VR input devices and head trackers. 15

16 Now that you see how EnSight can be used in a variety of ways on a variety of computers, how do you run everything? 16

17 Today s computer environments are numerous, complex, and typically difficult to access in a way that is as simple as just running an application on your desktop or laptop. How can EnSight be used in these computational environments and made as easy to use as a simple desktop application...such as EnSight just running on the desktop?! 17

18 CEIStart were added to EnSight starting with EnSight 9 as a prototype. EnSight 10.1 they are now easy to use for many users because of the redesign of CEIStart and the default configurations Remote Server and Remote SOS. These two configurations make starting EnSight easy for remote Server and remote SOS use. 18

19 More importantly, a customer can configure CEIStart to support the customer s computation environment. Some CEI customers have custom configurations for their various computational resources (e.g. different clusters and supercomputers). 19

20 20

21 Now that you have seen the many ways that EnSight can be used and the many choices you have for where to run the various EnSight components to address remote computing, distributed computing, and scaling EnSight for big data, we will now mention a few more tips for working with big data. 21

22 The default action is to load all parts. Don t do this if all parts aren t needed for postprocessing. Also, don t do this if you wish to change the element representation (next slide). 22

23 Nonvisual should be used for fluid domains or any large parts where you do not need to see the elements directly. Nonvisual means that the EnSight Client will not have any data for this part; only the Server has the data. This is very different from turning off the part s visibility. The Client still contains all the data for the part s geometry even if the part is invisible. Computing a part s feature angle representation can be very slow depending upon the number of elements and their type. Consider using a different element representation if you experience this. Feature Angle is the default representation for 3D data. 23

24 What is an engineer s hourly cost vs. the productivity gain? Modern graphics cards are very inexpensive when compared this way. 24

25 EnSight Flipbooks by default cache the visible polygon data on the EnSight Client. If this data is large or if you have a lot of time steps, then you can quickly use a lot of memory. If it takes to long to either load or play the EnSight flipbook, use: EnLiten, Reveal, or EnVideo instead. 25

26 EnSIght is a great application for analysis and visualizing data. EnVideo, EnLiten, and Reveal are designed for presentations, demos, or sharing data. Use the right application. 26

27 27

28 EnSight Minions is a Python tool that splits a long transient animation into several smaller transient animations. Each of the smaller transient animations can be created in parallel using multiple copies of EnSight. EnVe is then used automatically to join the several smaller movies into one large movie for the entire simulation. 28

29 29

30 30

31 31

32 EnSight command language is automatically created when you use EnSight. All actions are recorded to a file. When you quit EnSight you can save the command language file. You can also save the file explicitly. EnSight 10.0 did not always generate EnSight command language. EnSight 10.1 fixes this. EnSight also has Python command language support. Python is a real programming language. Use Python if you need a powerful language. You can record entire sessions using EnSight command language for simple tasks (such as toggle part visibility or add an annotation) or complex tasks (such as load multiple data sets, create parts, and load animations). Python can do the same but you must write the code. Python can be used with GUI code such as PyQt for creating custom user interfaces. At this time neither language can access the simulation data contained in the EnSight Server. However, query data and various summary information about the simulation data can be accessed. This tends to be much more powerful in Python. See the EnSight Interface Manual and Command Language Manual. 32

33 This slide and the next three show a custom user interface written in Python and PyQt used to post process STAR CCM+ data in EnSight and generate standard analysis output. The user never sees the normal EnSight user interface. 33

34 34

35 35

36 36

37 37