HighPerformance MemoryBased Web Servers: Kernel and UserSpace Performance

1
High-Performance Memory-Based Web ServersKernel
and User-Space Performance

• Philippe Joubert, Robert B. King, Rich Neves,
• Mark Russinovich, John M. Tracey
• USENIX Annual Technical Conference, 2001

2
Introduction

• This paper analyzes performance gap between
• User-mode caching web servers
• Kernel-mode web servers
• To identify potential performance gains possible
  for future user-mode primitives

3
Web Server Performance Issues

• Data copies and reads
• Event notification
• Queuing of a client request by a server for
  response by a server task
• TCP 3-way handshaking
• Receive data forming the request, read this data
  into user-mode memory area
• MP(multiple process/thread) - Apache
• SPED (single process event driven) Zeus, IIS
• Communication code path
• Socket layer - TCP/IP stack link layer - NIC

4
Current Approaches

• User-mode approaches
• To optimize existing interfaces and their
  implementations
• Modified select(), poll() to reduce event
  notification overhead
• To define completely new interfaces
• IO-Lite, WinNTs TransmitFile() API
• Kernel-mode approaches
• MS SWC(Scalable Web Cache), Linux kHTTPd,Linux
  TUX
• Other approaches
• Lava hit-server, Cheetah Web server (MIT
  Exokernel)

5
AFPA

• Adaptive Fast Path Architecture
• SW architecture for high-performance network
  servers
• AFPA features
• Support for various application protocols (HTTP,
  FTP, LDAP, DNS )
• Direct integration with TCP/IP protocol stack
• Kernel-managed zero copy cache
• Available with
• Linux, Windows 2000, AIX, OS/390

6
AFPA

• Data movement
• Data copies are avoided by passing reference to
  pinned cache object
• Reads are eliminated by avoiding checksum
  computation when sending cache objects as
  responses
• Event notification
• Scheduling and context switching overhead in
  responding to TCP/IP events is significantly
  reduced or eliminated
• Communication code path
• AFPA interfaces directly with TCP/IP (without
  socket layer)

7
Experimental Methodology

• Workload
• SPECWeb96
• Webstone
• Test environment
• Windows2000 Advanced Server
• AFPA on Windows 2000
• IIS
• SWC
• Red Hat 6.1 (2.3.51 kernel)
• AFPA on Linux
• kHTTPd
• TUX
• Zeus
• Apache

8
Experimental Methodology
9
Performance - SPECWeb96
10
Performance - Webstone (fixed size)
11
Performance - Webstone (fixed size)
Efficiency of zero copy TCP send
12
Conclusion

• Best performing user-mode web servers are at
  least 2 times slower than faster kernel-mode web
  server
• 3 attributes of best performing web servers
• Zero copy
• Efficient event notification mechanism
• Minimize communication code path
• Using new socket APIs
• Eliminating socket layer altogether