REMOTE RENDERING OF LARGE BIOLOGICAL DATASETS

Wolfram Schoor, R

udiger Mecke, Udo Seiffert

Fraunhofer Institute for Factory Operation and Automation (IFF), Sandtorstrasse 22, D-39106 Magdeburg, Germany

Felix Bollenbeck and Uwe Scholz

Leibniz Institute of Plant Genetics and Crop Plant Research, Corrensstr.3, D-06466 Gatersleben, Germany

Keywords:

Remote visualization, Scientiﬁc data, Hybrid rendering.

Abstract:

This paper presents remote rendering technologies to support the analysis and information exploration of large

biological datasets. Reconstructed three-dimensional models and other data domains stored in a database are

the foundation for the visualization. Remote Rendering via a clients web interface enables a widespread access

to the data with interactive frame rates and high-quality rendering results without the demand of new client

hardware. Moreover, the rendering server has the capability to manage multiple independent or collaborative

sessions and has a very small bandwidth consumption over time.

1 INTRODUCTION

Remote Rendering commonly means interactive vi-

sualization of three-dimensional datasets via a client/

server architecture over a network. Visualization of

complex 3D models is still a task hard to accomplish

without specialized hardware. Furthermore, ad-hoc

visualizations over distance are playing a key role in

future visualization trends. One problem of visual-

ization is that huge amounts of data must be handled,

which requires not only a fast CPU and high memory

capacities (RAM), but also a high degree of graphics

output generated by modern graphics cards.

These drawbacks can be overcome by the concept of

remote rendering. The datasets are computed on a

rendering machine with high graphics capabilities -

the so-called server, and only rendered images are

sent to the working station - called clients.

The work presented here is motivated by a given

shortcomming regarding biological model visualiza-

tion, namely limited bandwidth, slow clients with

poor graphics capabilities, large biological 3D mod-

els and the demand to real-time interaction. The lat-

ter cannot be covered by existing conventional remote

rendering techniques. Therefore, a remote rendering

architecture with reduced network trafﬁc generating

correct high-quality renderings for low level end user

PCs with interactive frame rates is given in this paper.

2 RELATED WORK

The idea of remote rendering is not new. Various re-

mote rendering applications have been developed in

the past. Adapted from (Schmalstieg and Gervautz,

1996), these techniques can be classiﬁed as follows:

Data

Visualization

Rendering

Display

Data

Visualization

Rendering

Data

Visualization

Rendering

Data

Visualization

Rendering

Display

Figure 1: Remote rendering types a) image-based, b) geom-

etry replication, c) immediate-mode, and d) hybrid.

Image-based. Rendering is performed by the sender,

and the resulting stream of pixels is sent over the

network as provided in (Scheiﬂer and Gettys, 1986)

(see Figure 1 (a)).

Geometry-based. A copy of the geometric database

223

Schoor W., Mecke R., Seiffert U., Bollenbeck F. and Scholz U.

REMOTE RENDERING OF LARGE BIOLOGICAL DATASETS.

DOI: 10.5220/0001802102230227

In Proceedings of the Fourth International Conference on Computer Graphics Theory and Applications (VISIGRAPP 2009), page

ISBN: 978-989-8111-67-8

is stored locally to be accessed by the rendering

process. The database can either be available be-

fore application start (kept on local hard disk), or

downloaded just before usage, such as VRML / X3D

browsers can do (Gueziec et al., 1999) (see Fig-

ure 1 (b)).

Immediate-mode Drawing. The low-level drawing

commands used by drawing APIs are issued by the

application performing the rendering. They are not

immediately executed, but sent over the network

as a kind of remote procedure call. The actual

rendering is then performed by the remote machine

(e.g. distributed GL (Shreiner et al., 2005)), see

Figure 1 (c).

Hybrid Rendering. In this approach a 3D geometry

as low resolution representation is rendered on the

client’s side and the server transmits images to

the client as well. The used mode depends on the

performed user action (computational vs. network

load) (Koller et al., 2004) or is dynamically adapted

to the available network capacities (see Figure 1 (d)).

Summarized, existing remote rendering tech-

niques cover a wide range of problem statements

including image-based remote renderings (suitable

for fast network connections) or geometry-based re-

mote renderings (e.g. cluster rendering with sufﬁ-

cient client’s graphics hardware). Hybrid rendering

methods for their capabilities could be identiﬁed as

adequate rendering method for low and medium level

clients using networks with medium network speed.

3 SYSTEM DEMANDS

In this section the requirements of the involved under-

lying system architecture and its components as well

as special user interests are discussed. Main issues

are the individual clients’ hardware conditions in the

network, the data transfer capabilities and the trait of

scientiﬁc data.

Clients Condition. A medium-sized scientiﬁc bio in-

stitution with approximately 300 potential clients was

evaluated. The premises of these clients are com-

pletely heterogeneous, they cover a big inventory of

out-of-date machines

as well as a few state-of-the-

art hardware machines. The present operating sys-

tems are as diverse as the clients’ hardware condition.

Whereas the processor speed is widely feasible, the

graphics hardware is not. These preconditions have

to be taken into account in the design of a remote vi-

sualization application for large biological models.

e.g. Win98 Pentium III with less or equal 500 MB RAM

Server Condition. Two different server machines

could be used for testing the remote visualization per-

formance.

The ﬁrst one is a high performance computer with

two Quad-Core Opteron 2.6 GHz, 16 GB RAM, a

Quadro FX 5600 graphics card and furthermore a

Tesla S870 processor with 500 GFlop and a shared

1 GBit connection (bottleneck) to the bio institution.

An average server upload B

of 2 MBit with average

ping rates t

avg

lesser than 23 ms could be identiﬁed

(σ

≈ 0.2 MBit, σ

avg

3 ms )

The second server is a minimal graphics server con-

ﬁguration with a 2.8 GHz Pentium IV with 4 GB

RAM. The rendering job is done by one Nvidia 7800

GTX graphics card. This server is directly located at

the bio institution. The minimum network connec-

tion for this setup is bound to the local switches and

achieves 100 MBit.

Data Transfer. The data transfer capabilities are one

major aspect in the remote data visualization and can

be formulated regarding to Equation 1.

B ≥

) C

depth

f n

f actor

(1)

Without any compression c

f actor

and for a screen

resolution R

(XGA), a color depth C

depth

3 Bytes/pixel and a frame rate f of 26 frames/sec for

only one client n, this would require at least a band-

width B of:

B ≥

(1024 ∗ 768) pixel ∗ 3 Bytes/pixel ∗ 26 sec

−1

∗ 1

B ≥ 60 MByte/s (2)

which is still high and only possible with at least

1 GBit/s ethernet

(see example calculation in Equa-

tion 2).

Using JPEG compression can reduce the image size

by a factor c

f actor

of 10 − 300 (depending on the re-

sulting image quality) and will also work with slower

connections. The local network speed at the bio in-

stitution depends on the used switches and includes

100 MBit as well as 1 GBit connections. This fact

prevents a standard remote rendering strategy due to

the reason that high-quality renderings would not ﬁt

into a 100 MBit connection if a couple of other users

share the same server connection (big n).

Data Resources. The biological data pool consists

of different data domains from biological material.

Available data include speciﬁcally microtome serial

section data, gene expression assays utilizing mRNA

average measured with http://speedtest.com

SXGA requires around 100 MByte/s.

GRAPP 2009 - International Conference on Computer Graphics Theory and Applications

224

in-situ hybridization (ISH) probing of histological

cross-sections as well as magnetic resonance spec-

troscopy measurements (

H-NMR) of caryopses at

different points in time. Other potential datasets, such

as macroarray gene expression data or metabolomic

assays by laser microdissection (LMD), could be inte-

grated into the database to enhance synergy and facili-

tate the inference of analysis results. Figure 2 presents

a pie chart breaking down the data distribution per

model.

Resources of data domains

24%

12%

60%

Histological data

3-D surface data

In-situ

hybridization (ISH)

H-NMR

Macroarray gene

expression

Metabolic assays

Figure 2: Data distribution for one model, the absolute data

totaling around 9 GB.

Visualization Aspects. Another very important pre-

condition is the function volume of the remote ren-

dering application to develop. At the moment the bio

researchers test a software application which is es-

pecially tailored to them, namely the BioVizIt appli-

cation. This software is initially a single-user stand

alone application, which is not network compatible

but covers the rest of the desired functionality.

All these features and actions have to be con-

sidered as part of the resulting remote rendering

application to best possibly support the biologist’s

work.

As conclusion, there are different possibilities to

perform remote visualization tasks as proposed in

Section 2. As possible candidates only those solutions

are appropriate which meet the requirements (recall

Section 3).

As solely remote rendering approach which ﬁts the

system requirements a hybrid rendering approach is

applicable. Due to the speciﬁc visualization require-

ments depicted in this section a conception with an

explicit control to data exchange mechanisms and

adaption possibilities of the visualization is necessary.

A solution to achieve this is to extend the existing

BioVizIt application with speciﬁc hybrid remote ren-

dering capabilities.

4 THE HYBRID REMOTE

RENDERING APPROACH

In the following, the remote visualization concept is

presented, where a hybrid rendering approach is the

basis. Furthermore, the architecture’s adaptions of the

BioVizIt platform, referring to network compatibility

and approaches addressing limited bandwidth, are de-

scribed.

Hybrid Rendering. The following Figure 3 illus-

trates the idea of the hybrid approach.

Rendered image

Client

Server

Datasets' names

(Scene graph changes)

Camera position and orientation

Rendering scene

Loading data sets

Creating session

Restoring session

Rendered image

(Scene graph changes)

Camera position and orientation

Resizing image

Compressing image

Rendering scene

Resizing image

Compressing image

Displaying image

Navigation

Modifying scene graph

Displaying image

Navigation

Modifying scene graph

Initial rendering request

Consecutive image requests

Image related Scene graph related

Viewpoint related

Session related

3D related

Reduced 3D dataset

Loading LOD model

Figure 3: Sequential diagram of the proposed hybrid ren-

dering approach.

In the beginning of the rendering process (initial ren-

dering request) a request with datasets’ names and

initial settings is sent to the server. As a result, the

server application is loading the requested ﬁle from

the database, and in addition a low resolution 3D ge-

ometry model (level of detail also LOD) is transferred

to the client. For optimization purposes the same ge-

ometry model is loaded only once. Moreover, the

graphics server submits the ﬁrst remote-rendered im-

age to the client.

If in the following, user navigation is performed (con-

secutive rendering requests), the local 3D model is

rendered by the client itself. Fast movements can be

achieved without having performance problems due

to network limitations. The local 3D model can be re-

duced up to one per mill of the original 3D model’s

data amount. Even if the bandwidth is less than

10 MBit an initial loading delay for an unoptimized

data transfer of 4 seconds for the complete low reso-

lution model would be acceptable to the user, if after-

wards a stable real-time interaction can be performed.

In case of stopping the navigation or changing the

scene graph for example, a command is immediately

sent to the server (containing at least the new camera

REMOTE RENDERING OF LARGE BIOLOGICAL DATASETS

225

position and orientation) to request a new rendered

image. This approach ensures real-time model inter-

actions with a high-quality rendering result during the

model exploration.

Geometry Submission: The server stores the informa-

tion of the original 3D models and also in coarser res-

olutions (LOD). Datasets are generated according to

the proposed technique in (Schoor et al., 2009).

If a dataset is requested, the original dataset is loaded

into the server application. At the same time a low

resolution model is transferred from the database to

the client.

Image Submission: The image submission by demand

is a concept to overcome the server network load. Ad-

ditionally, image compression algorithms JPEG2000

(Taubman and Marcellin, 2001) are used to reduce the

network trafﬁc.

Network Interconnect. Another important aspect

besides the underlying data transfer per se is how the

data will be transported. To achieve high resolution

rendering results on the client, only a reliable data

transfer method is an option. This excludes for ex-

ample UDP data transfer, because in this mode the

correct rendering result is not warranted due to un-

ordered or lost packages. The Transmission Control

Protocol TCP is sufﬁcient for the required purposes.

Adaption of the BioVizIt Platform. For an ex-

tension of the existing solution as server variant the

functionality must be appended to network and ses-

sion routines. Multiple network connections through

clients must be promptly accepted and processed, us-

ing multi-threading techniques. User events are reg-

istered in a message queue. This queue is held in

the operating system and members of the queue are

passed to the application (for example panning the

window). Received messages by the application are

then handled in a procedure of the primary thread. To

avoid blockages in the processing of this thread, a new

thread – the “server thread” – is generated during the

program start to wait for new client requests via TCP.

If a client connects to the server over a predeﬁned

port, a new “client thread” is generated immediately

for this client. This thread is responsible for further

communication with the client.

The client thread reads the request from the server’s

socket thread and interprets it through the HyperText

Transfer Protocol (HTTP). By accessing a virtual doc-

ument, additionally the document area of the HTTP

request is used, which includes the remote rendering

control commands of the client for the server. These

commands are processed sequentially.

At the end of this processing the client is enqueued to

a list, which contains clients awaiting an image syn-

thesis step.

5 RESULTS

In this section, the essential aspects of the implemen-

tation are presented. Key issues for remote visualiza-

tion are examined regarding the presented approach.

Latency of Network. The latency at the beginning of

a session depends on how the geometry transmission

is handled. The coarse 3D model can ﬁrst be accessed

after full model upload. This can take a few seconds

for the geometry transmission. This is an initial step

which is only performed once a model is loaded. Af-

ter that, the interaction is performed with latencies

< 30 ms due to local rendering. The server render-

ing result takes less than 1 s after the mouse button is

released (indicating the end of user navigation).

Graphics Performance. The graphics performance

of the client rendering stage depends on its hardware

and on the level of detail in which the 3D model was

submitted. For the tested application even on a Pen-

tium III client a coarse 3D model with 20,000 points

could be interactively visualized. During the obser-

vation of the model no rendering step is required (no

model interaction). That means that a real-time appli-

cation can be guaranteed for the model interaction.

Network Throughput. The network throughput per

high resolution image request (> 3-5 s) is equivalent

to the physical size of this image and corresponds to

approximately 300 kB. The geometry transfer per low

resolution 3D model is approximately 4-5 MB in size.

Scalability. The approach provided here has a good

scalability. The rendering of scenarios, such as de-

scribed in Section 3, can be accomplished on graphics

servers without problems with interactive frame rates

(Case 2: minimum graphics server solution). That

means, in the worst case

the remote rendering image

of the last handled client will be displayed with an ad-

ditional delay of approximately 1 s if the data transfer

is ignored. This is the theoretical maximum number

of users which can be handled by the server under the

assumption that a delay of more than 2 seconds is not

acceptable to users. Taking the 100 MBit connection

as lower bound for the data channel, more than 40

clients can request a full resolution image at the same

time without causing additional delay at a client sta-

tion. Assumed that a 2 MBit connection is the average

connection condition for the server case 1, and a full

image resolution is requested, only one client request

per second is possible without causing an additional

delay on the client.

Error Recovery. The error recovery is automatically

performed by the network protocol layer (TCP) which

provides explicit error corrections. Furthermore, a

e.g. a synchronous server request of 25 clients

GRAPP 2009 - International Conference on Computer Graphics Theory and Applications

226

transmission control and data ordering is also part of

this protocol.

Network Burst. In general, the failure of a network

connection in the beginning of a work session leads to

an insufﬁcient remote visualization. During the ses-

sion a temporary burstiness of the network connec-

tion in the time scale of seconds can be bridged with-

out problems due to the local downsampled version of

the 3D model. A loss of connectivity within a larger

time range disallows a practical use of the here pro-

vided technique due to the missing detail information

the remote-rendered image provides.

Security Issues. The data pool of plant biological

data consists of conﬁdential information. Even an

internal use must contain at least a simple protec-

tion mechanism due to project internal information.

Therefore, a simple login query was integrated to

authenticate the users with respective project rights.

Furthermore, each session on the remote visualization

server is initially hidden to other users.

Balance between Local/Remote Resources. The ar-

chitecture of the hardware at the scientiﬁc institution

is naturally very heterogeneous. A uniform strategy

to extensively use clients’ hardware is therefore not

applicable. As the least common denominator, a low

resolution 3D model, which is tailored to the actual

hardware conditions, is used at the clients’ side to

achieve real-time interaction.

Hardware/Software Incompatibility. The visual-

ization software is written in C++ and uses only

platform-independent libraries. The visualization ser-

vice tests for hardware premises, if speciﬁc features

are not supported by the server hardware, alterna-

tively CPU-based implementations exist (e.g. defor-

mation model). The client’s visualization front-end

uses a platform-independent web interface, based on

Java



Version 1.5, Java 3D

Figure 4 shows a screenshot of the provided web in-

terface and the visualization options.

Figure 4: Screenshot of the clients web interface showing a

reconstructed model of a caryopsis with clipping plane.

6 CONCLUSION AND OUTLOOK

In this paper a remote rendering approach was pre-

sented, which is applicable to medium-sized enter-

prises who have to deal with a large amount of sci-

entiﬁc data. With a minimum of effort and the use

of existing hardware this approach allows to visual-

ize, explore and modify data. The clients hardware

constitution must not be state-of-the-art which makes

this approach of visualization furthermore attractive

in a ﬁnancial manner. The network bandwidth could

be identiﬁed as limiting constraint respective to the

number of clients contemporaneously used, therefore

the local server with the minimal graphics conﬁgura-

tion (case 2) is to prefer.

If the remote rendering application will be used with

many users (> 30), the aid of more than one render-

ing machine can be useful. In the next term a larger

focus will be placed on data modiﬁcation possibili-

ties. Another point is the popping effect which occurs

by switching between the local LOD and the server

image. This effect is not so perturbing as occurring

in dynamic scenes. Anyway, a fading of the rendered

image can reduce this effect.

This research work is being supported by the BMBF

grants 0313821B and 0313821A.

REFERENCES

Anhøj, J. (2003). Generic design of web-based clinical

databases. Journal of Medical Internet Research,

5(4):e27.

Gueziec, A., Taubin, G., Horn, B., and Lazarus, F. (1999). A

Framework for Streaming Geometry in VRML. IEEE

Computer Graphics and Applications, 19(2):68–78.

Koller, D., Turitzin, M., Levoy, M., Tarini, M., Croccia, G.,

Cignoni, P., and Scopigno, R. (2004). Protected inter-

active 3d graphics via remote rendering. ACM Trans.

Graph., 23(3):695–703.

Scheiﬂer, R. W. and Gettys, J. (1986). The X window sys-

tem. ACM Trans. Graph., 5(2):79–109.

Schmalstieg, D. and Gervautz, M. (1996). Demand-driven

geometry transmission for distributed virtual environ-

ments. Computer Graphics Forum, 15(3):421–431.

Schoor, W., Bollenbeck, F., Weier, D., Weschke, W., Preim,

B., Seiffert, U., and Mecke, R. (2009). VR-Based Vi-

sualization and Exploration of Plant Biological Data.

JVRB. submitted.

Shreiner, D., Woo, M., Neider, J., and Davis, T. (2005).

OpenGL(R) Programming Guide: The Ofﬁcial Guide

to Learning OpenGL(R), Version 2 (5th Edition)

(OpenGL). Addison-Wesley Professional.

Taubman, D. S. and Marcellin, M. W. (2001). JPEG

2000: Image Compression Fundamentals, Standards

and Practice. Kluwer Academic Publishers, USA.

REMOTE RENDERING OF LARGE BIOLOGICAL DATASETS

227