projects: nfs-perf: jnewsome

I have moved my old page, which consisted mostly of raw benchmark results. The old page is here.

My area of interest is the RPC layer. I am working to analyze how the RPC currently functions, and to find ways to improve its performance. My individual test results are below.

Here is a walkthrough of the implementation of the RPC layer in the Linux 2.4.4 kernel. I will be referring to some of the key ideas presented here in the rest of my results.

Custom Tools

I have developed some tools to help analyze the performance and behavior of the RPC layer. Below are short descriptions of these tools, as well as links to their source code.

kprofile	Patch to the linux kernel that provides a profiling tool. Measurements are taken via rdtsc. Updated 1/25/2002
rpcstat	Patch to the linux kernel that provides information on current RPC activity via a proc interface. Also included is a program that reads the proc files at a given interval in order to monitor how the RPC behaves over time.

Test Results

Below are results from some of the tests I've taken.

Baseline Read/Write Tests	Simple measurement of read and write performance on files from 20 MB to 2000 MB, using the Sun Connectathon basic5 test. UPDATED 7/30/01
Network Latency Measurements	Measurement of network latency during a write test. Network activity was captured with tcpdump and parsed by ethereal.
RPC Analysis with rpcstat	Analysis of a 2000 MB write to our Network Appliance F85 file server.
Follow-up Analysis with rpcstat	Analysis of a 2000 MB write to a Linux NFS server, and a comparison to the rpcstat results from the F85.
Follow-up Analysis with rpcstat	Another analysis of a 2000 MB write to our F85 server. This time a larger RAID array was used, which fixed the problems I was seeing before.
Analysis with kprofile	Analysis of results taken with kprofile, a kernel profiling tool which is available on our tools page. The purpose is to show where the RPC layer is spending its time.