Experimental Results

 

We ran our experiments on four Intel and Alpha hardware configurations:

• Pentium-II/440LX. These are Dell Dimension XPS D-300 workstations containing a 300 MHz Intel Pentium-II processor and an Intel 440LX chipset. Each machine has 128MB of RAM and a Myricom Lanai 4.1 SAN adapter (M2M-PCI32C) connected to a 32-bit 33 MHz PCI slot.
• Pentium-II/440BX. These machines use a Pentium-II processor clocked at 450 MHz on an Asus P2B motherboard with an Intel 440BX chipset. Each machine has 128MB of RAM and a Myricom Lanai 4.1 LAN adapter (M2F-PCI32) connected to a 32-bit 33 MHz PCI slot.

  
  

    

  Figure 2: TCP Bandwidth

  
  
  
• DEC Miata. These are Digital Personal Workstation 500au platforms with a 500MHz Alpha 21164 CPU, a 96KB L2 cache, a 2MB L3 cache, and the Digital 21174 ``Pyxis'' chipset. These machines are configured with 512MB of RAM and a Myricom Lanai 4.1 SAN adapter connected to a 32-bit 33 MHz PCI slot. The Pyxis limits I/O bandwidth to approximately 103 MB/s on the sending side.
• DEC Monet. These are Compaq XP1000 Professional Workstations, with a 500 MHz Alpha 21264 CPU, a 4MB L2 cache, and the Digital 21272 ``Tsunami'' chipset. These machines are configured with 640MB of RAM and a Myricom Lanai 5.2 SAN adapter connected to a 64-bit 33 MHz PCI slot. The Lanai-5 is described in http://www.myri.com:80/scs/PCI64X/PCI64X-spec.html. We also report some Gigabit Ethernet measurements from this platform, using Alteon ACENIC adapters based on the Tigon-II chipset (firmware revision 12.3.8), interconnected through an ACEswitch 1080 (firmware revision 5.0.24).

All systems run kernels built from the same FreeBSD 4.0 source pool, which was current as of 4/15/99. The hosts are interconnected through diverse Myrinet switch models, which have no measurable effects on the results.

To take timings, we used netperf version 2.1pl3 built from the FreeBSD ports collection. We modified netperf to collect CPU utilization by reading the system timers directly from kernel memory via libkvm, in order to correctly charge interrupt overhead to the netperf process. All tests were run on isolated machines, and the vast majority of the interrupts serviced came from the gigabit NIC.