Automatic Parameter Configuration Mechanism for Data Transfer Protocol GridFTP

1
Automatic Parameter Configuration Mechanism for
Data Transfer Protocol GridFTP

• Takeshi ItoHiroyuki OhsakiMakoto Imase
• Graduate School of Information Science and
  TechnologyOsaka University, Japan

2
Contents

• Background
• GridFTP
• Automatic Parameter Configuration Mechanism
  (APCM)
• Design strategy
• Basic idea
• Simulation
• Conclusions

3
Background

• TCP (Transmission Control Protocol)
• Has been widely used in the Internet
• Has a difficulty in maintaining high throughput
• In a network with high packet loss probability
• GridFTP
• A protocol to effectively transfer large volume
  data
• Solve the existing TCP problems
• Currently under standardization by GGF (The
  Global Grid Forum)

4
GridFTP

• Features
• Parallel data transfer
• Automatic negotiation of TCP socket buffer size
• Third-party control of file transfer
• Partial file transfer
• Security
• Reliable file transfer
• Examination has not been conducted sufficiently
  so far
• How to configure GridFTP control parameters

5
Objectives

• Propose a mechanism
• i.e., Automatic Parameter Configuration Mechanism
  (APCM)
• Optimizes the number of parallel TCP connections
• Improves GridFTP performance
• Demonstrate the effectiveness of APCM

6
Related Works
9 S. Thulasidasan, W. Feng, and M. K. Gardner,
Optimizing GridFTP through dynamic
right-sizing, in Proceedings of IEEE
International Symposium on High Performance
Distributed Computing, June 2003.1 T. Ito, H.
Ohsaki, and M. Imase, On parameter tuning of
data transfer protocol GridFTP in wide-area Grid
computing, in Proceedings of Second
International Workshop on Networks for Grid
Applications (GridNets 2005), Oct. 2005, pp.
415.421.

• Proposed a method 9
• Determine the required TCP socket buffer size
• Realized by extending the SBUF command
• Optimal number of parallel TCP connections and
  TCP socket buffer size are derived 1
• Using a TCP fluid-flow model
• Required information should be known in advance
• Round-trip time, available bandwidth

7
Parallel Data Transfer

• Higher throughput can be expected
• In comparison to using a single TCP connection
• Throughput drops if the number of TCP connections
  becomes too large

A single file can be transferred
throughmultiple TCP connections
8
Design Strategy

• It is extremely important to
• Provide the compatibility with existing GridFTP
• Realize APCM on the side of the GridFTP client
• Establish APCM without changing the existing
  GridFTP protocol
• Realize APCM in the Grid middleware
• It is difficult to realize APCM in the
  third-party transfer
• Most of the transfers are executed between
  GridFTP servers and clients
• Not be so problematic

9
Basic Idea
data channel
control channel
2. Transfer a chunk of the file to be transferred
1. Fix the number of TCP connections
a chunk
Grid Network
file
GridFTP client
GridFTP server
4. Update the number of TCP connections,
and another chunk is transferred
3. Measure the network status on the GridFTP
client
10
Measuring Network Status
data channel
control channel

• Measure on the side of GridFTP client

GridFTP goodputCalculate from the size of
chunkand its transfer time
Round-trip timeCommand response timein the
GridFTP control channel
Grid Network
file
GridFTP client
GridFTP server
TCP socket buffer sizeUse the socket API
TCP socket buffer sizeUse a feature of GridFTP
(SBUF)
11
Adjusting the Number of Parallel TCP Connections

• APCM adjusts the number of parallel TCP
  connections
• Discuss three operational modes
• MI (Multiplicative Increase)
• MI (Multiplicative Increase Plus)
• AIMD (Additive Increase and Multiplicative
  Decrease)

12
MI (Multiplicative Increase) Mode
Number of parallel TCP connections
If (4) is not satisfied, terminate the algorithm
If (4) is satisfied, increase the number
ofparallel TCP connections
Initialize the number ofparallel TCP connections
k-th chunk transfer
1 2 3 4
13
MI (Multiplicative Increase Plus) Mode
If (7) is not satisfied, numerically derivethe
number of parallel TCP connections,which
maximize Eq. (1) 1
Number of parallel TCP connections
If (7) is satisfied, increase the number
ofparallel TCP connections
Configure the number ofparallel TCP
connections to that value, andterminate the
algorithm
Initialize the number ofparallel TCP connections
k-th chunk transfer
1 2 3 4
14
AIMD (Additive Increase and Multiplicative
Decrease) Mode
Number of parallel TCP connections
If (10) is not satisfied, decrease the number
ofparallel TCP connections
If (10) is satisfied, increase the number
ofparallel TCP connections

Initialize the number ofparallel TCP connections
k-th chunk transfer
1 2 3 4 5 6 7
15
Network Topology Used in Simulation
chunk size 100 MbyteTCP socket buffer size
64 Kbyte
A file of 10Gbyte istransferred from the
GridFTP clientto the GridFTP server
A GridFTP server and a GridFTP clientare
connected via two DropTail routers
16
Simulation Result(Evolution of GridFTP Goodput
and Number of Parallel TCP connections)
The rise of the number of parallelTCP
connections is slowerin AIMD mode compared to
bothMI mode and MI mode
By using MI and MI mode,GridFTP can
sufficiently useavailable bandwidth
17
Simulation Result (Effect of the Bottleneck Link
Bandwidth)

• For comparison purposes, GridFTP goodput with a
  single TCP connection is also included

By using APCM, the bottleneck linkbandwidth can
bealmost fully utilized
Poor performance whenthe bottleneck link
bandwidth is large
18
Simulation Result (Effect of Propagation Delay)
With MI mode and MI mode, GridFTP goodput hardly
degrades even fora large propagation delay
With AIMD mode, GridFTP goodputdegrades
linearly asthe propagation delaybecomes large
19
Conclusions

• Proposed APCM for GridFTP
• Focusing on the parallel data transfer feature
• Transfer a file as chunks and measure GridFTP
  goodput and round-trip time
• Three operational modes (MI, MI, and AIMD)
• Evaluated effectiveness by simulation experiments
• Significantly improve GridFTP goodput

20
Future Works

• Conduct simulation experiments
• Under various network configurations
• Identify a parameter range that optimizes the
  performance of APCM
• Evaluate fairness among GridFTP sessions
• GridFTP sessions share the same bottleneck link

21
Thank you for your kind attention
22
Why is Design Strategy important? (1/2)

• Provide the compatibility with existing GridFTP
• GridFTP is implemented in the Globus Toolkit
• GridFTP has been spreading rapidly in recent
  years
• Realize APCM on the side of the GridFTP client
• A number of GridFTP servers has been in operation
• Establish APCM without changing the existing
  GridFTP protocol
• To interconnect with existing GridFTP servers

23
Why important? (2/2)

• Realize APCM in the Grid middleware
• APCM can be easily installed in Grid computing
  environment
• APCM needs to operate in various computer
  environments as well as various network
  environments
• Grid computing is featured by the heterogeneity
  of computers and networks constituning Grid

24
What means the equation (4)?
GridFTP goodput is upper-bound by N W / Rwhen
the number of connections is too small
? is a parameter (0 lt ? lt 1)to consider effect
of slow-start phase of TCP
So, if the inequality (4) is satisfied,conjecture
the TCP socket is bottleneck
25
How numerically derive the number of parallel TCP
connections from Eq. (1)?

• From Eqs. (1) and (2), G can be expressed as a
  function of N, W, R, B
• Can measure G, N, W, R and not B
• Can calculate B from Eqs. (1) and (2)
• Derive N which maximize G in the Eq. (1) by using
  the value of W, R, B

26
Why do you think of 3 modes?

• We dont know which mode performs better
• In a environment, MI mode may perform better
• In another environment, MI mode may perform
  better
• So, we think of 3 modes, and evaluate their
  effectiveness
• More through investigation of all 3 modes is
  required
• Under various network conditions

27
MI mode always performs better than MI mode?

• No, we dont think so.
• The analysis 1 is based on some assumptions
• When a assumption is not satisfied, the number of
  TCP connections could be misconfigured with MI
  mode
• Then, MI mode could perform better than MI mode
• More investigation under various network
  configurations is needed

28
AIMD mode is not useful?

• We think AIMD mode is also useful.
• For example, in a general network with multiple
  GridFTP sessions
• Regarding fairness among GridFTP sessions, it is
  expected that AIMD mode may perform better than
  MI and MI mode

29
Why utilization of the bottleneck link bandwidth
is slightly degraded as the bottleneck link
bandwidth becomes large?

• APCM takes time to find the optimal number of
  parallel TCP connections
• TCP connections take time to fully utilize the
  bottleneck link bandwidth when the bottleneck
  link bandwidth is large

30
Why not measure the packet loss probability and
available bandwidth?

• It is difficult to measure at the middleware
  layer using a passive measurement method
• Should features specific to certain operating
  systems or network devices on the computers be
  used, it would become possible to measure
• However, ease of installation and/or deployment
  in the Grid computing environment should be
  compromised
• Even if could measure by using active measurement
• Compatibility with existing GridFTP servers would
  be impaired