Internet Engineering Task Force (IETF) V. Singh
Request for Comments:
8868 callstats.io
Category: Informational J. Ott
ISSN: 2070-1721 Technical University of Munich
S. Holmer
Google
January 2021
Evaluating Congestion Control for Interactive Real-Time Media
Abstract
The Real-Time Transport Protocol (RTP) is used to transmit media in
telephony and video conferencing applications. This document
describes the guidelines to evaluate new congestion control
algorithms for interactive point-to-point real-time media.
Status of This Memo
This document is not an Internet Standards Track specification; it is
published for informational purposes.
This document is a product of the Internet Engineering Task Force
(IETF). It represents the consensus of the IETF community. It has
received public review and has been approved for publication by the
Internet Engineering Steering Group (IESG). Not all documents
approved by the IESG are candidates for any level of Internet
Standard; see
Section 2 of RFC 7841.
Information about the current status of this document, any errata,
and how to provide feedback on it may be obtained at
https://www.rfc-editor.org/info/rfc8868.
Copyright Notice
Copyright (c) 2021 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(
https://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with respect
to this document. Code Components extracted from this document must
include Simplified BSD License text as described in Section 4.e of
the Trust Legal Provisions and are provided without warranty as
described in the Simplified BSD License.
Table of Contents
1. Introduction
2. Terminology
3. Metrics
3.1. RTP Log Format
4. List of Network Parameters
4.1. One-Way Propagation Delay
4.2. End-to-End Loss
4.3. Drop-Tail Router Queue Length
4.4. Loss Generation Model
4.5. Jitter Models
4.5.1. Random Bounded PDV (RBPDV)
4.5.2. Approximately Random Subject to No-Reordering Bounded
PDV (NR-BPDV)
4.5.3. Recommended Distribution
5. Traffic Models
5.1. TCP Traffic Model
5.2. RTP Video Model
5.3. Background UDP
6. Security Considerations
7. IANA Considerations
8. References
8.1. Normative References
8.2. Informative References
Contributors
Acknowledgments
Authors' Addresses
1. Introduction
This memo describes the guidelines to help with evaluating new
congestion control algorithms for interactive point-to-point real-
time media. The requirements for the congestion control algorithm
are outlined in [
RFC8836]. This document builds upon previous work
at the IETF: Specifying New Congestion Control Algorithms [
RFC5033]
and Metrics for the Evaluation of Congestion Control Algorithms
[
RFC5166].
The guidelines proposed in the document are intended to help prevent
a congestion collapse, to promote fair capacity usage, and to
optimize the media flow's throughput. Furthermore, the proposed
congestion control algorithms are expected to operate within the
envelope of the circuit breakers defined in
RFC 8083 [
RFC8083].
This document only provides the broad set of network parameters and
traffic models for evaluating a new congestion control algorithm.
The minimal requirement for congestion control proposals is to
produce or present results for the test scenarios described in
[
RFC8867] (Basic Test Cases), which also defines the specifics for
the test cases. Additionally, proponents may produce evaluation
results for the wireless test scenarios [
RFC8869].
This document does not cover application-specific implications of
congestion control algorithms and how those could be evaluated.
Therefore, no quality metrics are defined for performance evaluation;
quality metrics and the algorithms to infer those vary between media
types. Metrics and algorithms to assess, e.g., the quality of
experience, evolve continuously so that determining suitable choices
is left for future work. However, there is consensus that each
congestion control algorithm should be able to show that it is useful
for interactive video by performing analysis using real codecs and
video sequences and state-of-the-art quality metrics.
Beyond optimizing individual metrics, real-time applications may have
further options to trade off performance, e.g., across multiple
media; refer to the RMCAT requirements [
RFC8836] document. Such
trade-offs may be defined in the future.
2. Terminology
The terminology defined in RTP [
RFC3550], RTP Profile for Audio and
Video Conferences with Minimal Control [
RFC3551], RTCP Extended
Report (XR) [
RFC3611], Extended RTP Profile for RTCP-Based Feedback
(RTP/AVPF) [
RFC4585] and Support for Reduced-Size RTCP [
RFC5506]
applies.
3. Metrics
This document specifies testing criteria for evaluating congestion
control algorithms for RTP media flows. Proposed algorithms are to
prove their performance by means of simulation and/or emulation
experiments for all the cases described.
Each experiment is expected to log every incoming and outgoing packet
(the RTP logging format is described in
Section 3.1). The logging
can be done inside the application or at the endpoints using PCAP
(packet capture, e.g., tcpdump [tcpdump], Wireshark [wireshark]).
The following metrics are calculated based on the information in the
packet logs:
1. Sending rate, receiver rate, goodput (measured at 200ms
intervals)
2. Packets sent, packets received
3. Bytes sent, bytes received
4. Packet delay
5. Packets lost, packets discarded (from the playout or de-jitter
buffer)
6. If using retransmission or FEC: post-repair loss
7. Self-fairness and fairness with respect to cross traffic:
Experiments testing a given congestion control proposal must
report on relative ratios of the average throughput (measured at
coarser time intervals) obtained by each RTP media stream. In
the presence of background cross-traffic such as TCP, the report
must also include the relative ratio between average throughput
of RTP media streams and cross-traffic streams.
During static periods of a test (i.e., when bottleneck bandwidth
is constant and no arrival/departure of streams), these reports
on relative ratios serve as an indicator of how fairly the RTP
streams share bandwidth amongst themselves and against cross-
traffic streams. The throughput measurement interval should be
set at a few values (for example, at 1 s, 5 s, and 20 s) in
order to measure fairness across different timescales.
As a general guideline, the relative ratio between congestion-
controlled RTP flows with the same priority level and similar
path RTT should be bounded between 0.333 and 3. For example,
see the test scenarios described in [
RFC8867].
8. Convergence time: The time taken to reach a stable rate at
startup, after the available link capacity changes, or when new
flows get added to the bottleneck link.
9. Instability or oscillation in the sending rate: The frequency or
number of instances when the sending rate oscillates between an
high watermark level and a low watermark level, or vice-versa in
a defined time window. For example, the watermarks can be set
at 4x interval: 500 Kbps, 2 Mbps, and a time window of 500 ms.
10. Bandwidth utilization, defined as the ratio of the instantaneous
sending rate to the instantaneous bottleneck capacity: This
metric is useful only when a congestion-controlled RTP flow is
by itself or is competing with similar cross-traffic.
Note that the above metrics are all objective application-independent
metrics. Refer to
Section 3 of [netvc-testing] for objective metrics
for evaluating codecs.
From the logs, the statistical measures (min, max, mean, standard
deviation, and variance) for the whole duration or any specific part
of the session can be calculated. Also the metrics (sending rate,
receiver rate, goodput, latency) can be visualized in graphs as
variation over time; the measurements in the plot are at one-second
intervals. Additionally, from the logs, it is possible to plot the
histogram or cumulative distribution function (CDF) of packet delay.
3.1. RTP Log Format
Having a common log format simplifies running analyses across
different measurement setups and comparing their results.
Send or receive timestamp (Unix): <int>.<int> -- sec.usec decimal
RTP payload type <int> -- decimal
SSRC <int> -- hexadecimal
RTP sequence no <int> -- decimal
RTP timestamp <int> -- decimal
marker bit 0|1 -- character
Payload size <int> -- # bytes, decimal
Each line of the log file should be terminated with CRLF, CR, or LF
characters. Empty lines are disregarded.
If the congestion control implements retransmissions or Forward Error
Correction (FEC), the evaluation should report both packet loss
(before applying error resilience) and residual packet loss (after
applying error resilience).
These data should suffice to compute the media-encoding independent
metrics described above. Use of a common log will allow simplified
post-processing and analysis across different implementations.
4. List of Network Parameters
The implementors are encouraged to choose evaluation settings from
the following values initially:
4.1. One-Way Propagation Delay
Experiments are expected to verify that the congestion control is
able to work across a broad range of path characteristics, including
challenging situations, for example, over transcontinental and/or
satellite links. Tests thus account for the following different
latencies:
1. Very low latency: 0-1 ms
2. Low latency: 50 ms
3. High latency: 150 ms
4. Extreme latency: 300 ms
4.2. End-to-End Loss
Many paths in the Internet today are largely lossless; however, in
scenarios featuring interference in wireless networks, sending to and
receiving from remote regions, or high/fast mobility, media flows may
exhibit substantial packet loss. This variety needs to be reflected
appropriately by the tests.
To model a wide range of lossy links, the experiments can choose one
of the following loss rates; the fractional loss is the ratio of
packets lost and packets sent:
1. no loss: 0%
2. 1%
3. 5%
4. 10%
5. 20%
4.3. Drop-Tail Router Queue Length
Routers should be configured to use drop-tail queues in the
experiments due to their (still) prevalent nature. Experimentation
with Active Queue Management (AQM) schemes is encouraged but not
mandatory.
The router queue length is measured as the time taken to drain the
FIFO queue. It has been noted in various discussions that the queue
length in the currently deployed Internet varies significantly.
While the core backbone network has very short queue length, the home
gateways usually have larger queue length. Those various queue
lengths can be categorized in the following way:
1. QoS-aware (or short): 70 ms
2. Nominal: 300-500 ms
3. Buffer-bloated: 1000-2000 ms
Here the size of the queue is measured in bytes or packets. To
convert the queue length measured in seconds to queue length in
bytes:
QueueSize (in bytes) = QueueSize (in sec) x Throughput (in bps)/8
4.4. Loss Generation Model
Many models for generating packet loss are available: some generate
correlated packet losses, others generate independent packet losses.
In addition, packet losses can also be extracted from packet traces.
As a (simple) minimum loss model with minimal parameterization (i.e.,
the loss rate), independent random losses must be used in the
evaluation.
It is known that independent loss models may reflect reality poorly,
and hence more sophisticated loss models could be considered.
Suitable models for correlated losses include the Gilbert-Elliot
model [gilbert-elliott] and models that generate losses by modeling a
queue with its (different) drop behaviors.
4.5. Jitter Models
This section defines jitter models for the purposes of this document.
When jitter is to be applied to both the congestion-controlled RTP
flow and any competing flow (such as a TCP competing flow), the
competing flow will use the jitter definition below that does not
allow for reordering of packets on the competing flow (see NR-BPDV
definition below).
Jitter is an overloaded term in communications. It is typically used
to refer to the variation of a metric (e.g., delay) with respect to
some reference metric (e.g., average delay or minimum delay). For
example in
RFC 3550, jitter is computed as the smoothed difference in
packet arrival times relative to their respective expected arrival
times, which is particularly meaningful if the underlying packet
delay variation was caused by a Gaussian random process.
Because jitter is an overloaded term, we use the term Packet Delay
Variation (PDV) instead to describe the variation of delay of
individual packets in the same sense as the IETF IP Performance
Metrics (IPPM) working group has defined PDV in their documents
(e.g.,
RFC 3393) and as the ITU-T SG16 has defined IP Packet Delay
Variation (IPDV) in their documents (e.g., Y.1540).
Most PDV distributions in packet network systems are one-sided
distributions, the measurement of which with a finite number of
measurement samples results in one-sided histograms. In the usual
packet network transport case, there is typically one packet that
transited the network with the minimum delay; a (large) number of
packets transit the network within some (smaller) positive variation
from this minimum delay, and a (small) number of the packets transit
the network with delays higher than the median or average transit
time (these are outliers). Although infrequent, outliers can cause
significant deleterious operation in adaptive systems and should be
considered in rate adaptation designs for RTP congestion control.
In this section we define two different bounded PDV characteristics,
1) Random Bounded PDV and 2) Approximately Random Subject to No-
Reordering Bounded PDV.
The former, 1) Random Bounded PDV, is presented for information only,
while the latter, 2) Approximately Random Subject to No-Reordering
Bounded PDV, must be used in the evaluation.
4.5.1. Random Bounded PDV (RBPDV)
The RBPDV probability distribution function (PDF) is specified to be
of some mathematically describable function that includes some
practical minimum and maximum discrete values suitable for testing.
For example, the minimum value, x_min, might be specified as the
minimum transit time packet, and the maximum value, x_max, might be
defined to be two standard deviations higher than the mean.
Since we are typically interested in the distribution relative to the
mean delay packet, we define the zero mean PDV sample, z(n), to be
z(n) = x(n) - x_mean, where x(n) is a sample of the RBPDV random
variable x and x_mean is the mean of x.
We assume here that s(n) is the original source time of packet n and
the post-jitter induced emission time, j(n), for packet n is:
j(n) = {[z(n) + x_mean] + s(n)}.
It follows that the separation in the post-jitter time of packets n
and n+1 is {[s(n+1)-s(n)] - [z(n)-z(n+1)]}. Since the first term is
always a positive quantity, we note that packet reordering at the
receiver is possible whenever the second term is greater than the
first. Said another way, whenever the difference in possible zero
mean PDV sample delays (i.e., [x_max-x_min]) exceeds the inter-
departure time of any two sent packets, we have the possibility of
packet reordering.
There are important use cases in real networks where packets can
become reordered, such as in load-balancing topologies and during
route changes. However, for the vast majority of cases, there is no
packet reordering because most of the time packets follow the same
path. Due to this, if a packet becomes overly delayed, the packets
after it on that flow are also delayed. This is especially true for
mobile wireless links where there are per-flow queues prior to base
station scheduling. Owing to this important use case, we define
another PDV profile similar to the above, but one that does not allow
for reordering within a flow.
4.5.2. Approximately Random Subject to No-Reordering Bounded PDV (NR-
BPDV)
No Reordering BPDV, NR-BPDV, is defined similarly to the above with
one important exception. Let serial(n) be defined as the
serialization delay of packet n at the lowest bottleneck link rate
(or other appropriate rate) in a given test. Then we produce all the
post-jitter values for j(n) for n = 1, 2, ... N, where N is the
length of the source sequence s to be offset. The exception can be
stated as follows: We revisit all j(n) beginning from index n=2, and
if j(n) is determined to be less than [j(n-1)+serial(n-1)], we
redefine j(n) to be equal to [j(n-1)+serial(n-1)] and continue for
all remaining n (i.e., n = 3, 4, .. N). This models the case where
the packet n is sent immediately after packet (n-1) at the bottleneck
link rate. Although this is generally the theoretical minimum in
that it assumes that no other packets from other flows are in between
packet n and n+1 at the bottleneck link, it is a reasonable
assumption for per-flow queuing.
We note that this assumption holds for some important exception
cases, such as packets immediately following outliers. There are a
multitude of software-controlled elements common on end-to-end
Internet paths (such as firewalls, application-layer gateways, and
other middleboxes) that stop processing packets while servicing other
functions (e.g., garbage collection). Often these devices do not
drop packets, but rather queue them for later processing and cause
many of the outliers. Thus NR-BPDV models this particular use case
(assuming serial(n+1) is defined appropriately for the device causing
the outlier) and is believed to be important for adaptation
development for congestion-controlled RTP streams.
4.5.3. Recommended Distribution
Whether Random Bounded PDV or Approximately Random Subject to No-
Reordering Bounded PDV, it is recommended that z(n) is distributed
according to a truncated Gaussian for the above jitter models:
z(n) ~ |max(min(N(0, std^(2)), N_STD * std), -N_STD * std)|
where N(0, std^(2)) is the Gaussian distribution with zero mean and
std is standard deviation. Recommended values:
std = 5 ms
N_STD = 3
5. Traffic Models
5.1. TCP Traffic Model
Long-lived TCP flows will download data throughout the session and
are expected to have infinite amount of data to send or receive.
This roughly applies, for example, when downloading software
distributions.
Each short TCP flow is modeled as a sequence of file downloads
interleaved with idle periods. Not all short TCP flows start at the
same time, i.e., some start in the ON state while others start in the
OFF state.
The short TCP flows can be modeled as follows: 30 connections start
simultaneously fetching small (30-50 KB) amounts of data, evenly
distributed. This covers the case where the short TCP flows are
fetching web page resources rather than video files.
The idle period between bursts of starting a group of TCP flows is
typically derived from an exponential distribution with the mean
value of 10 seconds.
| These values were picked based on the data available at
| <
https://httparchive.org/reports/state-of-the- | web?start=2015_10_01&end=2015_11_01&view=list> as of October
| 2015.
Many different TCP congestion control schemes are deployed today.
Therefore, experimentation with a range of different schemes,
especially including CUBIC [
RFC8312], is encouraged. Experiments
must document in detail which congestion control schemes they tested
against and which parameters were used.
5.2. RTP Video Model
[
RFC8593] describes two types of video traffic models for evaluating
candidate algorithms for RTP congestion control. The first model
statistically characterizes the behavior of a video encoder, whereas
the second model uses video traces.
Sample video test sequences are available at [xiph-seq]. The
following two video streams are the recommended minimum for testing:
Foreman (CIF sequence) and FourPeople (720p); both come as raw video
data to be encoded dynamically. As these video sequences are short
(300 and 600 frames, respectively), they shall be stitched together
repeatedly until the desired length is reached.
5.3. Background UDP
Background UDP flow is modeled as a constant bit rate (CBR) flow. It
will download data at a particular CBR for the complete session, or
will change to particular CBR at predefined intervals. The inter-
packet interval is calculated based on the CBR and the packet size
(typically set to the path MTU size, the default value can be 1500
bytes).
Note that new transport protocols such as QUIC may use UDP; however,
due to their congestion control algorithms, they will exhibit
behavior conceptually similar in nature to TCP flows above and can
thus be subsumed by the above, including the division into short-
lived and long-lived flows. As QUIC evolves independently of TCP
congestion control algorithms, its future congestion control should
be considered as competing traffic as appropriate.
6. Security Considerations
This document specifies evaluation criteria and parameters for
assessing and comparing the performance of congestion control
protocols and algorithms for real-time communication. This memo
itself is thus not subject to security considerations but the
protocols and algorithms evaluated may be. In particular, successful
operation under all tests defined in this document may suffice for a
comparative evaluation but must not be interpreted that the protocol
is free of risks when deployed on the Internet as briefly described
in the following by example.
Such evaluations are expected to be carried out in controlled
environments for limited numbers of parallel flows. As such, these
evaluations are by definition limited and will not be able to
systematically consider possible interactions or very large groups of
communicating nodes under all possible circumstances, so that careful
protocol design is advised to avoid incidentally contributing traffic
that could lead to unstable networks, e.g., (local) congestion
collapse.
This specification focuses on assessing the regular operation of the
protocols and algorithms under consideration. It does not suggest
checks against malicious use of the protocols -- by the sender, the
receiver, or intermediate parties, e.g., through faked, dropped,
replicated, or modified congestion signals. It is up to the protocol
specifications themselves to ensure that authenticity, integrity,
and/or plausibility of received signals are checked, and the
appropriate actions (or non-actions) are taken.
7. IANA Considerations
This document has no IANA actions.
8. References
8.1. Normative References
[
RFC3550] Schulzrinne, H., Casner, S., Frederick, R., and V.
Jacobson, "RTP: A Transport Protocol for Real-Time
Applications", STD 64,
RFC 3550, DOI 10.17487/
RFC3550,
July 2003, <
https://www.rfc-editor.org/info/rfc3550>.
[
RFC3551] Schulzrinne, H. and S. Casner, "RTP Profile for Audio and
Video Conferences with Minimal Control", STD 65,
RFC 3551,
DOI 10.17487/
RFC3551, July 2003,
<
https://www.rfc-editor.org/info/rfc3551>.
[
RFC3611] Friedman, T., Ed., Caceres, R., Ed., and A. Clark, Ed.,
"RTP Control Protocol Extended Reports (RTCP XR)",
RFC 3611, DOI 10.17487/
RFC3611, November 2003,
<
https://www.rfc-editor.org/info/rfc3611>.
[
RFC4585] Ott, J., Wenger, S., Sato, N., Burmeister, C., and J. Rey,
"Extended RTP Profile for Real-time Transport Control
Protocol (RTCP)-Based Feedback (RTP/AVPF)",
RFC 4585,
DOI 10.17487/
RFC4585, July 2006,
<
https://www.rfc-editor.org/info/rfc4585>.
[
RFC5506] Johansson, I. and M. Westerlund, "Support for Reduced-Size
Real-Time Transport Control Protocol (RTCP): Opportunities
and Consequences",
RFC 5506, DOI 10.17487/
RFC5506, April
2009, <
https://www.rfc-editor.org/info/rfc5506>.
[
RFC8083] Perkins, C. and V. Singh, "Multimedia Congestion Control:
Circuit Breakers for Unicast RTP Sessions",
RFC 8083,
DOI 10.17487/
RFC8083, March 2017,
<
https://www.rfc-editor.org/info/rfc8083>.
[
RFC8593] Zhu, X., Mena, S., and Z. Sarker, "Video Traffic Models
for RTP Congestion Control Evaluations",
RFC 8593,
DOI 10.17487/
RFC8593, May 2019,
<
https://www.rfc-editor.org/info/rfc8593>.
[
RFC8836] Jesup, R. and Z. Sarker, Ed., "Congestion Control
Requirements for Interactive Real-Time Media",
RFC 8836,
DOI 10.17487/
RFC8836, January 2021,
<
https://www.rfc-editor.org/info/rfc8836>.
8.2. Informative References
[gilbert-elliott]
Hasslinger, G. and O. Hohlfeld, "The Gilbert-Elliott Model
for Packet Loss in Real Time Services on the Internet",
14th GI/ITG Conference - Measurement, Modelling and
Evalutation [sic] of Computer and Communication Systems,
March 2008,
<
https://ieeexplore.ieee.org/document/5755057>.
[netvc-testing]
Daede, T., Norkin, A., and I. Brailovskiy, "Video Codec
Testing and Quality Measurement", Work in Progress,
Internet-Draft, draft-ietf-netvc-testing-09, 31 January
2020,
<
https://tools.ietf.org/html/draft-ietf-netvc-testing-09>.
[
RFC5033] Floyd, S. and M. Allman, "Specifying New Congestion
Control Algorithms", BCP 133,
RFC 5033,
DOI 10.17487/
RFC5033, August 2007,
<
https://www.rfc-editor.org/info/rfc5033>.
[
RFC5166] Floyd, S., Ed., "Metrics for the Evaluation of Congestion
Control Mechanisms",
RFC 5166, DOI 10.17487/
RFC5166, March
2008, <
https://www.rfc-editor.org/info/rfc5166>.
[
RFC8312] Rhee, I., Xu, L., Ha, S., Zimmermann, A., Eggert, L., and
R. Scheffenegger, "CUBIC for Fast Long-Distance Networks",
RFC 8312, DOI 10.17487/
RFC8312, February 2018,
<
https://www.rfc-editor.org/info/rfc8312>.
[
RFC8867] Sarker, Z., Singh, V., Zhu, X., and M. Ramalho, "Test
Cases for Evaluating Congestion Control for Interactive
Real-Time Media",
RFC 8867, DOI 10.17487/
RFC8867, January
2021, <
https://www.rfc-editor.org/info/rfc8867>.
[
RFC8869] Sarker, Z., Zhu, X., and J. Fu, "Evaluation Test Cases for
Interactive Real-Time Media over Wireless Networks",
RFC 8869, DOI 10.17487/
RFC8869, January 2021,
<
https://www.rfc-editor.org/info/rfc8869>.
[tcpdump] "Homepage of tcpdump and libpcap",
<
https://www.tcpdump.org/index.html>.
[wireshark]
"Homepage of Wireshark", <
https://www.wireshark.org>.
[xiph-seq] Daede, T., "Video Test Media Set",
<
https://media.xiph.org/video/derf/>.
Contributors
The content and concepts within this document are a product of the
discussion carried out in the Design Team.
Michael Ramalho provided the text for the jitter models
(
Section 4.5).
Acknowledgments
Much of this document is derived from previous work on congestion
control at the IETF.
The authors would like to thank Harald Alvestrand, Anna Brunstrom,
Luca De Cicco, Wesley Eddy, Lars Eggert, Kevin Gross, Vinayak Hegde,
Randell Jesup, Mirja Kühlewind, Karen Nielsen, Piers O'Hanlon, Colin
Perkins, Michael Ramalho, Zaheduzzaman Sarker, Timothy B. Terriberry,
Michael Welzl, Mo Zanaty, and Xiaoqing Zhu for providing valuable
feedback on draft versions of this document. Additionally, thanks to
the participants of the Design Team for their comments and discussion
related to the evaluation criteria.
Authors' Addresses
Varun Singh
CALLSTATS I/O Oy
Rauhankatu 11 C
FI-00100 Helsinki
Finland
Email: varun.singh@iki.fi
URI:
https://www.callstats.io/ Jörg Ott
Technical University of Munich
Department of Informatics
Chair of Connected Mobility
Boltzmannstrasse 3
85748 Garching
Germany
Email: ott@in.tum.de
Stefan Holmer
Google
Kungsbron 2
SE-11122 Stockholm
Sweden