CMSC 838B : Information Visualization

Application Project : Disk I/O Access Pattern

Yoo Ah Kim

Data set : Disk I/O trace from HP lab

Visualization Tool : Spotfire 5.1

Data Set Description

• Type 
  • Read/Write 
  • Asynchronous/Synchronous 
  • Block type: user data, inode, directory, indirect block, cylinder group, etc. 
• Timimgs 
  • Enqueue time: when the disk driver first sees the request 
  • Start time: when the request is sent to the disk 
  • Stop time: when the request returns from the disk 
• Disk number 
• Start address 
• Transfer size
  

The traced system is a single-user workstation. The main uses of the system were electronic mail and editing the papers. Due to the file buffer cache, there was not much disk activity on this system. The following two tables show the detail of the system.

Table 1. The Computer System Traced

Table 2. The File System Configuration

Analysis of Data

Figure 1. Hourly Value for a Week 

Timings

Figure 2-1 and 2-2 show the queueing delay distribution for read and write. This is caused by system load and I/O burstiness. We can find out that write operations are more bursty, therefore cause more delay.

Figure 2-1. Start Time Distribution - Read

Figure 2-2. Start Time Distribution - Write

Figure 3 shows processing time distribution. Most of processing times are between 10,000us - 50, 000us. I excluded some erroneous results.

Figure 3. Processing Time Distribution

Types

As we can see in figure 4-2, user data percentage is only 27% in I/O access and the remaining parts are metadata, especially inode access.

Figure 4-1. Read/Write                                                                             Figure 4-2. Types of Blocks

The HP-UX file system generates both synchronous and asynchronous requests. Most of reads and half of writes are synchronous.

Figure 4-3. Sync/Async

Sequential Access

This data is disk level I/O traces, not file system level. Due to UNIX buffer cache, many of events don't arrive at disk level at all. In a result, traces are not really sequential, especially read operations. There are small areas accessed sequentially.

Figure 5. Enqueue time vs. Address

I/O Sizes 

Most of I/O sizes are multiple of 2k(2k, 4k, 8k), but there are some writes for a large amount of blocks.

Figure 6. I/O Sizes (Read/Write)

Critique

• Spotfire is easy to learn and use
• Gives good overview of data
• Easy to find features by choosing fields and ranges dynamically
• Good to detect and leave out errors in dataset
• With more than  40, 000 records, 
  • Spotfire works fairly well
  • but it takes significant time (about 2 min) to import data
  • can't distinguish each data due to occlusion
  • 3D graph: a little slow and difficult to understand 
• limited choice of graph types

Written by Yoo Ah Kim  --  Feb. 26, 2001