Transcript Filers - Usenix

Surfing Technology Curves
Steve Kleiman
CTO
Network Appliance Inc.
1
Book Plug
The Innovator’s Dilemma - When New
Technologies Cause Great Firms to Fail
 Clayton M. Christensen
2
About NetApp
Two product lines:
 Network Attached File Servers (a.k.a. filers)
 Web proxy caches: NetCache
Founded in 1992
>$1B revenue run rate
>70% CAGR since founding
>120% last year
3
Filers: Fast, Simple, Reliable and Multi-protocol
System
Sun E 3500/4500
HP-9000 N4000
NetApp 840
CPUs
2
4
1
Overall Resp. Ops per
Result Resp. @ Max SpecRate
8,165
15,270
15,235
3.04
1.91
1.54
23.8
3.7
3.6
20.4
10.4
46
Ops/FS
RAID
340
318
15,235
no
yes
yes
4
Filers: Fast, Simple, Reliable and Multi-protocol
Disk management
 Filer finds disks and organizes into RAID groups
and spares automatically
 Simple addition of storage
 Automatic RAID reconstruction
Data management
 Snapshots
 SnapRestore
 SnapMirror
Simple upgrade
Small command set
5
Filers: Fast, Simple, Reliable and Multi-protocol
Built-in RAID
Easy hardware maintenance
 Hot plug disk, power, fans
 Low MTTR
Cluster Failover
Autosupport
>99.995% measured field availability
6
Filers: Fast, Simple, Reliable and Multi-protocol
NFS
CIFS
 CIFS and NFS attributes
HTTP
FTP
DAFS
Internet Cache
 FTP
 Streaming media
7
Wave 1:
Networks, Appliances and Software
8
Network and Storage Bandwidth
Year
1992
1994
1996
1998
2001
Storage
10 MB
20 MB
40 MB
100 MB
200-400 MB
Network
0.1 MB
1 MB
10 MB
100 MB
1000 MB
Penalty
100-to-1
20-to-1
4-to-1
1-to-1
.2-to-1
9
The Appliance Revolution
1980s
(General Purpose)
1990s
(Appliance Based)
UNIX
UNIX
Application
Windows/NT
Print
File Service
Filer
Routing
Router/Switch
Printer
...
...
10
Appliance philosophy
Appliance philosophy breeds focus
 External simplicity  internal simplicity
 RISC argument
Don’t have to be all things to all people
 Limited compatibility constraints
 Interfaces are bits on wire
 Think different!
Can innovate with both software and hardware
11
Filer Architecture
Commercial off-the shelf chips
 Any appropriate architecture
CPU
 i486  Pentium  Alpha ‘064
 Alpha‘164  PIII
Board level integration
 1 or more CPUs (4)
 1 or more PCI busses (4)
 High bandwidth switches
 Multiple memory banks
 Integrated I/O
NVRAM
Mem
PCI
NVRAM
12
Roads Not Taken
No “unobtainium”
 Minimalist infrastructure
 No special purpose busses
 No big MPs
 Motherboards only: no cache coherent
backplanes
No functionally distributed computers
No special purpose networks (e.g. HIPPI)
No block access protocols
13
DataOnTap Architecture
Daemons, Shells, Commands
Java Virtual Machine
Lib
ATM
GbE
FDDI
100BT
VI
NIC*
NFS
FCAL
TCP/IP
CIFS
WAFL
RAID
Disk
HTTP
SCSI
VIPL
DAFS
SK
* VI supported on FC, (Future: GbE, Infiniband)
14
DataOnTap
Simple Kernel
 Message passing
 Non-preemptive
Sample optimizations
 Checksum caching
 Suspend/Resume
 Cache hit pass through
15
WAFL: Write Anywhere File Layout
Log-like write throughput
 No segment cleaning (LFS)
 Write data allocated to optimize RAID performance
 Delayed write allocation
Active data is never overwritten (shadow paging)
 On-disk data is always consistent
 File system state is changed atomically
 Every 10 sec, by default
Client modification requests are logged to NVRAM
 NVRAM log is replayed only on reboot
16
Wave 2:
Memory-to-Memory Interconnects
(a.k.a NUMA, NORMA)
17
Problem:
Remove single points of failure
 Without doubling hardware
 Minimizing performance overhead
 Without decreasing reliability
18
Clustered Failover Architecture
Network
Filer 1
ServerNet
Filer 2
NVRAM
NVRAM
Fibre Channel
Fibre Channel
19
Memory-to-Memory Interconnects
Efficient transfer model
 Allows minimal overhead on receiver
Scaleable Bandwidth
 High speed ASIC based switching
 Gigabit technology
Open architecture
 PCI, not coherent bus interface
 Incorporate multiple technologies
Relatively inexpensive
20
Mirroring NVRAM
 NVRAM is split into
local and partner
regions
PCI Bus
ServerNet
DMA
 Data is assembled in
NVRAM
 Data is DMAed from
NVRAM to equivalent
offset in remote node
To partner
NVRAM
NVRAM data
from partner
CPU
NVRAM
 Client reply is sent
when log entry DMA
completes
21
Leveraged Components
Memory-to-Memory interconnects
 Low overhead, high-bandwidth, cheap
WAFL
 Always consistent file system
 Built-in NVRAM logging/replay
Fibre Channel disks
 Two independent ports
Single function appliance software
 Simple, low-overhead failover
22
Wave 3:
The Internet
23
The Consequences of
Higher-speed Internet Access
200K-400K home cable head-end
 Requires 1.5-3Gbps access capability
 30% subscription rate, 20% online
 Minimum 128Kbps BW
Enterprise
 Remote sites still connected by slow links
 Require high-quality access to content
 Overloaded web servers
ISP
 Require distribution and caching of large
media files
24
Yet Another Appliance
Cisco
NetApp
25
NetCache
HTTP/FTP proxy cache appliance
 Highly deployable
 Forward and reverse proxy
Transparency
Filtering
iCAP
 Enables value added services
 Virus scanning, transcoding, ad insertion, …
Stream splitting
Stream caching
Content distribution
26
Cacheable Content
Cacheable
Content
Time
Static
Content
Dynamic
Content
Streaming
Media
27
Wave 4:
The Death of Tapes
28
Using Tapes for Disaster Recovery
Year
Drive
#
Tapes
# Tape drives to
Capacity
Capacity Drives
Required restore in 8 hours
1999
36G
168
6TB
172
21
2000
72G
216
16TB
160
28
2001
144G
500a
72TB
360
63
a: with SAN
29
SnapMirror
Remote asynchronous mirroring
 Continuous incremental update
 Only allocated blocks are transmitted
 Automatic resynchronization after
disconnect
 Destination is always a consistent
“snapshot” of source
Filer
WAN
Filer
30
Creating a Snapshot
Before
Snapshot
After
Snapshot
Active FS
Active FS
Snapshot
A
B
C
After
Block Update
D
Disk Blocks
A
B
Active FS
Snapshot
C
D
A
B
C
D
C’
New
Block
31
WAFL: Block Map File
Multiple bits per 4KB block
 Column for allocated block
in the active file system
 Columns for allocated
blocks in snapshots
Taking a Snapshot
 Copy root inode
S1 S2 S3 FS
Block 1
Block 2
Block 3
Block 4
Block 5
Block 6
Block 7
Block 8
32
Consistent Image Propagation
 Fast Network or Slow Modification Rate
Source
1
2
3
Destination
4
1
5
4
6
5
6
 Slow Network or High Modification Rate
Source
Destination
1
2
3
4
1
5
6
4
33
Wave 5:
Local File Sharing and
Virtual Interface Architecture
34
ISPs: Scalable Services
Scalability
 Scale compute power and storage independently
Resiliency
Cost
 Commodity hardware and Open Systems standards
Load
Balancing
Switch
Application
Servers
Gigabit
Switch
File
Servers
F760
F760
Internet or
Intranet
F760
Data Center
35
Database
Better Manageability
 Offline backup with snapshots
 Replication
 Recovery from snapshots
 Easy storage management
Equal or better performance
 Less retuning
F760
36
Local File Sharing
Geographically constrained
 1 or 2 machine rooms
Mostly homogeneous clients
 Can be large or small
 1 - 100 machines
Single administrative control
High performance applications
 Web service, Cache
 Email, News
 Database, GIS
37
Local File Sharing Architecture Characteristics
Applications tend to avoid OS
 e.g. No virtual memory
Applications tend to have OS adaptation layer
Different access protocol requirements
 e.g. high-performance locking, recovery,
streaming
38
What is VI?
Virtual Interface (VI) Architecture
 VI architecture organization
 Promoted by Intel, Compaq and Microsoft
 VI Developer’s Forum
Standard capabilities
 Send/receive message, remote DMA read/write
 Multiple channels with send/completion queues
 Data transfer bypasses kernel
 Memory pre-registration
39
VI Architecture
Application
VIPL
Library
User
Kernel
KVIPL client
Data
Kernel
Control
KVIPL
Module
Hardware
VI compliant
NIC driver
VI compliant
NIC
40
VI-compliant implementations
Fibre channel (FC-VI draft standard)
 e.g. Troika, Emulex
Giganet
Servernet II
Infiniband
 Enables 1U MP heads
Future: VI over TCP/IP
41
How VI Improves Data Transfer
No fragmentation, reassembly and
realignment data copies
No user/kernel boundary crossing
No user/kernel data copies
 Data transfer direct to application buffers
42
Direct Access File System
Application
Buffers
User
File Access API
DAFS
VIPL* API
VIPL
VI NIC
Driver
Kernel
Data
Control
Hardware
NIC
Memory
* VI Provider Layer specification
maintained by the VI Developers Forum
43
DAFS Benefits
 File access protocol with implicit data sharing
 Direct application access
 File data transfers directly to application buffers
 Bypasses Operating System
 File semantics
 Optimized for high throughput and low latency
 Consistent high speed locking
 Graceful recovery/failover of clients and servers
 Fencing
 Enhanced data recovery
 Leverages VI for transport independence
44
DAFS vs. SAN
Wires
Direct
(direct transfer to memory)
Block
Local
Attached
Network
(TCP/IP)
SCSI over IP
SAN
Protocols
File
DAFS
NAS
45
Summary
Wave 1: Filers
 Technology: Fast networks, commodity
servers
 Environment: Appliance-ization
Wave 2: Failover
 Technology: Memory-to-memory
interconnects, Dual ported FC disks
 Environment: 24x7 requirements
Wave 3: NetCache
 Technology: Internet, HTTP
 Environment: High BW requirements, POP
deployability
46
Summary
Wave 4: SnapMirror
 Technology: Disk areal density, Fibre
Channel, fast networks
 Environment: Cost of downtime for recovery
Wave 5: DAFS
 Technology: VI architecture
 Environment: Local file sharing
47