Measuring disk throughput

CSL373 : Operating Systems

This homework is based on the ZCAV tool: see ZCAV documentation here.

Background

Given constant rotational speed (constant angular velocity, CAV) of a magnetic disk, the head can cover larger distance on the outer tracks in the same time, compared to the inner tracks. If the sector density on all tracks is same, this would cause space wastage on the outer tracks (as their density would be limited by the inner tracks).

To deal with this situation, modern disks divide their tracks into zones. The inner zones have lower angular sector density, while the outer zones have higher angular sector density. Effectively, this means that the disk throughput at the outer zones is higher. This is also called zoned constant angular velocity (ZCAV).

To take advantage of this, the disk controllers usually map the lower sector numbers (starting from sector number zero) on the outer tracks, while the higher sector numbers are mapped to inner tracks. The thinking being that the lower sector numbers are likely to have higher frequency of access. Notice that this is just an assumption and may not be true!

ZCAV program measures the disk throughput at different sector numbers. To do this, it reads/writes 256MB (default) worth of data starting at sector X, where X is varied from 0 to disk-size.

You can see the results of running ZCAV on some disk configurations here.

Turn in

Run ZCAV using the following command for the following types of disks:

A disk on a physical desktop in GCL. We will provide you with root access on one of the desktops to perform this experiment.
A disk on a physical laptop (optional). If you have a laptop, report its make and model, and published disk characteristics (seek, latency, throughput). Then measure it using ZCAV and report. Also, report any interesting conclustions (if any). This part is optional; if you do not have access to a laptop, you are not required to do this part.
A USB (pen) drive. Report the make, model, and published characteristics of the USB drive and compare with your obtained results.
A disk mounted inside a Virtual Machine (VM). We ran the experiment on a Baadal VM ourselves and here is the raw data we obtained for two disks (main disk containing the OS image of 80GB size, and an attached disk of 8GB size): main_80GB.txt, attached_8GB.txt. You can answer this part by either interpreting this raw data, or generating data for yourself. Here are the credentials required to access the Baadal VM:
```
  IP address: 172.16.3.16
  user: root
  password: zcav123
  
```
To access this VM, you will first have to login to baadalcsegateway.cse.iitd.ernet.in using your GCL credentials, and then SSH to 172.16.3.16 from there.
The VM has 8 CPUs, and 8GB RAM, and a qcow2-format mounted disk. You may want to read a bit about how VMs are implemented (in particular, how disks are virtualized and how the qcow2 format works) to answer this question.

Report all results as graphs as at this URL. Feel free to use the GNUPlot scripts available at this URL (e.g., command.gnuplot.zcav). In particular, how many zones do you find? What is the mapping between sector numbers and physical disk layout? What are the corresponding disk bandwidths? Anything else you notice? Finally, explain the different behaviour in different environments.

Tips

Downloading and Installing: download this tarball, and type the following commands:
- ```
tar xzf bonnie++-1.03e.tgz
```
- ```
cd bonnie++-1.03e
```
- ```
./configure
```
- ```
make
```
Running: sudo ./zcav /dev/sda (replace sda with sdb, hda, etc. depending on the disk being tested).
You only need to test the read performance (not write).
You may first want to setup the experiment on your private machine before experimenting on GCL. Note that one experiment can take hours to complete, so plan in advance on accessing the GCL machine.
While zcav is running, you may want to check that your CPU is not bottlenecked (i.e., utilization is below 100\%) using top. Also, you may want to check that no other IO-intensive operation is running using iotop.

Fun

This part is optional. If you do these, feel free to submit your results.

Can you modify the ZCAV program to estimate the seek and latency of the magnetic disk? How? What were your results?
Try the same experiments for writes (instead of reads). Do your results differ? For which environments? Why? Warning: writes will overwrite your filesystem and can potentially make the FS unusable. So any write experiments should not be done on the disk containing the current FS.
Try the same experiments on filesystems (instead of raw disks as above). Here, you would need to create a large file, and then use the command:./zcav filename.
- Try for different filesystems and different FS options: in particular, try for ext4 ordered mode vs. ext4 default mode.