tnc 2007 bandwidth-on-demand to reach the optimal throughput of media brecht vermeulen stijn...
TRANSCRIPT
TNC 2007
Bandwidth-on-demand to reach the optimal throughput of media
Brecht Vermeulen
Stijn Eeckhaut, Stijn De Smet, Bruno Volckaert, Joachim Vermeir, Filip De Turck, Piet Demeester (Ghent University – IBBT)Ibrahim Habib, Zhaoming Li (City University of New York )
With acknowledgment to IBBT FIPA & GEISHA project, VRT, IBM & University of Antwerp
Brecht VermeulenTNC 2007
p. 2
Broadcaster workflow
Archiving
Media production
Raw videomaterial
ingest
ingest
Playout
Brecht VermeulenTNC 2007
p. 3
Tape based workflow
Digital tapes
Linear editing
Rough cut
Final editing
Voice over
Playout
Non-Linear editing
Local conversion to file
Brecht VermeulenTNC 2007
p. 4
File based workflow
Central storage
Playout
NLE clients
Archive
Digital files
Digital tape from camera
or memory device
Windows or Apple
Brecht VermeulenTNC 2007
p. 5
Files: so ?
SD
DV25 DNxHD 2.2 HDCAM HDCAM-SR Proxy
Mb/s 28,95 220 144 440 1,51h video (in GB) 13,0 99,0 64,8 198,0 0,7Transfer 1h video in seconds over Gb/s 111 843 551 1685 6
HD
Research issues: Optimal large file transfer: network & server
performance Offsite transcoding/rendering farms (& editing &
voice-over & subtitles, ...) File-based archiving Disaster recovery
Brecht VermeulenTNC 2007
p. 6
Contents
Introduction Optimising server networking
TCP/IP offloading vs. CPU based FTP vs. NFS vs. CIFS
Network based vision Ongoing research Conclusion
Brecht VermeulenTNC 2007
p. 7
TCP tuning options
Adapt kernel TCP parameters (free) Bigger receive window: more data in-transit Important if bandwidth*delay is high Linux: rmem,wmem,tcp_rmem, tcp_wmem,
mem,netdev_max_backlog Windows registry: Tcp1323Opts=3,
GlobalMaxTcpWindowSize,TcpWindowSize,AFD DefaultReceive(Send)window
e.g. Buffers and window on 4MB Jumbo frames ($)
MTU 9000 bytes, ..., 16000 bytes Not really a standard-> NICs, switches to be tested
Brecht VermeulenTNC 2007
p. 8
TCP offloading
TCP checksum & segmentation offload ($) Most modern good nics Works with standard kernel Warning: some cards say that they do offloading,
but it is done in the driver software Full TCP offload ($$)
Complete TCP/IP stack on the NIC (incl. retransmits, slow start...)
TCP setup/teardown still by host Webserver short connections vs. long transfers
Problems with e.g. Bonding Kernel patch needed (linux)
Brecht VermeulenTNC 2007
p. 9
TCP offloading
Normal NIC Offloading
Brecht VermeulenTNC 2007
p. 10
TCP offloading tests
Back-to-back tests between AMD dual Opteron systems (Opteron 246 @ 2GHz)
Intel PRO/1000 NIC (4 x 1 Gbps) TCP checksum & segm offload
Chelsio T204 TOE (4 x 1 Gbps) full TCP offload (= TCP Offload Engine)
TCP throughput measured with Iperf Generates TCP streams on different interfaces Transfers are memory-to-memory
Limitations PCI-X bus: 64 bit @ 133 MHz ~ 1GB/s
PCI-X is a half-duplex bus, PCI Express is a full-duplex point-to-point connection
Maximal (unidir) TCP efficiency: 94.1% 941 Mbps per link 99% for 9000 byte MTU
Brecht VermeulenTNC 2007
p. 11
Chelsio TOE vs. Intel Pro 1000 (MTU 1500) 4 links unidir: 3.7 Gb/s vs. Intel NIC 2.7 Gb/s 4 links bidir: 7 Gb/s vs. Intel NIC 3.2 Gb/s
Jumbo frames on Intel: throughput +, CPU -
TCP offloading results
Inte
l
Inte
lC
hels
io
Chels
io
MTU 1500
MTU 9000
4 Gb/s
8 Gb/s
50%
100%
Brecht VermeulenTNC 2007
p. 12
Protocol comparison setup
Transfers between storage and memory GPFS fibre channel storage used
360MB/s write, 690MB/s read from one server 2.88Gb/s write, 5.52 Gb/s read
Brecht VermeulenTNC 2007
p. 13
Protocol comparison
FTP > NFS > CIFS for reads FTP > CIFS > NFS for writes FTP with chelsio close to GPFS performance
Brecht VermeulenTNC 2007
p. 14
CIFS (synchr.) vs. Latency: model
Brecht VermeulenTNC 2007
p. 15
Contents
Introduction Optimising server networking Network based vision
Broadcasters’ problems Media grid farms Archiving Disaster recovery
Ongoing research Conclusion
Brecht VermeulenTNC 2007
p. 16
Broadcasters’ problems
Typically broadcasters work together with production houses, remote studios, ...
Storage and computing is not core business of broadcasters -> outsource to datacenters ?
Networking seems THE solution, BUT... FTP > NFS > CIFS+delay issue: but remember
windows editing clients -> CIFS HDCAM-SR: 440Mb/s video codec Storage bandwidth: both for archiving (and retrieve
something from archive), disaster recovery Time-critical (journals)
Brecht VermeulenTNC 2007
p. 17
Mediagrid farms
Editing on standard definition, rendering on rendering farms on HD (editing effects, cuts, ...)
Problems: Standard grid infrastructure is more directed
towards computing intensive vs. storage/dataset intensive tasks
For broadcasters: guarantees are needed on bandwidth and computing availability
Bandwidth to the rendering farms should be high, but can be by reservation (e.g. for non-live productions).
Brecht VermeulenTNC 2007
p. 18
Archive: (S)ATA disk price evolution
19992000 2001
20022003
20042005
20062007
Price per GB
0,0
1,0
2,0
3,0
4,0
5,0
6,0
7,0
8,0
9,0
10,0
Source: own purchase prices 1999-2007
Max Capacity €/GB
SATA 2000 75 7,4SCSI 2000 73 17,4SATA 2007 750 0,3SCSI 2007 300 2,6
Brecht VermeulenTNC 2007
p. 19
Archive
Cheaper disks and tape library systems: Online/nearline file-based archiving
Storage management
Brecht VermeulenTNC 2007
p. 20
Archive providers
Brecht VermeulenTNC 2007
p. 21
Archive needs
Only high bandwidth when retrieving content
Uploading of content may be slower Some content may be duplicated to two
sites, other to only one site Reservations for guaranteed bandwidth ?
Brecht VermeulenTNC 2007
p. 22
Disaster recovery
Central storage
Playout
NLE clients
Archive
Central storage is large Production is done on this Total restore = > 24 hours Solution:
Working on remote copy ? Networking/server performance ? Client CIFS ? Bandwidth guarantees on-demand for this ?
Brecht VermeulenTNC 2007
p. 23
Contents
Introduction Optimising server networking Network based vision Ongoing research
VPN between Gent and New York
Conclusion
Brecht VermeulenTNC 2007
p. 24
VPN: Gent – New York
For now: only 100Mb/s
Cuny (NY)
HOPINew YorkNew York
BELNETAmsterdamIBBT
Ingress LSP in NY Egress LSP in NL
HOPI-Ghent_Nyc_Ams
Transit LSP in BELNET’s GhentEgress LSP in GEANT2 NL
HOPI-Ghent_Ghent_AmsAmsterdam
New York
Ghent
Ingress LSP in GEANT2 NL
HOPI-Ghent_Ams_Ghent
LDP session to exchange vlan 4003
Ingress LSP in NLEgress LSP in NY
HOPI-Ghent_Ams_Nyc
Figure provided by Dante
Brecht VermeulenTNC 2007
p. 25
CVLSR
CHEETAH Virtual Label Switching Router Linux control PC with GMPLS engine Ethernet switch with
bandwidth reservations
Due to delay in setup and performance issues, research is still ongoing
One possible way
Brecht VermeulenTNC 2007
p. 26
Conclusions
Demand from broadcasters: Bandwidth and remote storage/computing Large files
Research: Optimal configuration and tuning of protocol
parameters and servers to use the bandwidth Is bandwidth reservation a solution for network
distribution of this functionality ?