Running Timer Code

From Earlham Cluster Department

Revision as of 06:38, 22 November 2005 by Tmcnulty (Talk | contribs)
Jump to: navigation, search


Obtaining Kernel Timing Data

Testing is now radically simplified with the script.

Setup a client and server with the tp_timer kernel

If you already have a client and server running the proper kernel, skip to the next section. tp_timer should be compiled in module format (simply copy the -mod config below, as indicated).

1. Build and install the kernel. To compile the kernel, check out llk-sources from CVS and use a .config file from llk-sources/ccg-configs. Make sure that CONFIG_TP_TIMER=y (for tp_timer built in) or CONFIG_TP_TIMER=m (for a module version of tp_timer). Note that there is a significant chance that this compilation will crash the build host is running tp_timer (having the module unloaded is okay). This is probably a problem with NFS, so building it locally may work (but watch out for distcc and ccache that like to touch files in your home directory). In short, build on a machine that isn't running a tp_timer kernel.

cvs co llk-sources
cd llk-sources
cp ccg-configs/<machine>-mod .config # where machine is 'cairo' or 'bazaar'
make oldconfig
make -jN CC="ccache gcc" # where N is the number of hosts in DISTCC_HOSTS * 4
./ # ikl = install kernel locally

2. Setup the boot-loader

# on cairo:
vim /etc/yaboot.conf
ybin -v

# on bazaar:
vim /boot/grub/menu.lst
grub < /etc/grub.conf

3. Reboot.


Check out and run the test script

This must be done on a non-testing node, e.g., hopper. If you run this on a node you're using to test and accidently touch NFS, the machine will crash.

cvs co llk
cd llk/scripts
./ <client> <server> <protocol> <table name> <test set>

Where <client> and <server> are short host names like 'b10' and the protocol is 'udp' or 'tcp'. The script will load the module on both hosts, calibrate the test, run the test, fetch the data, unload the module, and insert the data into the database with the specified table name and test set label.

Older info

Basic Flow:

  1. make sure the tp_timer module is installed and working on the client and server nodes
  2. clear /proc/tp_timer on both client and server machines (just read its contents; cat /proc/tp_timer > /dev/null)
  3. run pacgen on client and server to amass new data.
  4. read data into log files on both server and client
  5. load data from the log files into database
  6. look at data via php scripts

Machines setup with llk timing: b10,b11,c10,c11

Need to check out from CVS:

pacgen sample commands:

./server tcp <port number>
./client -t <protocol> -s <source IP> -m <source port> -d <dest IP> -p <dest port>

Each window is a specific ID for the packet. This ID is used as the payload for the packets. Useful for checking for packet fragmentation (educated guess via Alex). Windows are 0xffffffff, 0xffffffff (2 unsigned ints worth of 1's), an unsigned int for socket/thread id (We don't use it; that part of the code was removed), and an unsigned int holding the window number (we don't know what this is for).

After pacgen is ran with on a tp timer enabled machine, the timing results will be available in /proc/tp_timer. Note that after you read this proc entry, the information will be cleared.

To enter the data into our database:

  1. Collect the /proc/tp_timer from the server node and the client node:
    1. cat /proc/tp_timer > tp_timer_client.log
  2. Execute the script found in the the llk cvs module in the scripts sub directory.
    1. Usage: ./ <tablename> <label> <datafile> <client|server>
    2. If the table given by name to the scripts doesn't exist, the script creates the table.
Personal tools
this semester