tune(1) · carta

NAME

tune – calibrate MOSIX kernel parameters to suit processor and network

SYNOPSIS

tune

mtune

tune_kernel

prep_tune

DESCRIPTION

tune measures the timing of certain kernel operations in MOSIX, resulting in 14 tuning parameters. Similarly, mtune measures the timing of MFS operations, producing 6 other tuning parameters.

Although the MOSIX kernel will run reasonably well with the pre-supplied parameters, optimal results are achieved by measuring different parameters for each set of nodes whose hardware differs by one or more of the following:

1.: the processor
2.: the type or speed of the memory
3.: the number of memory wait-states (if any)
4.: the networking hardware
5.: whether the kernel has the SMP option configured or not

tune produces two output files, corresponding to two different methods of installing the parameters: The first output file, named dscosts.h, is to be included as is in the kernel sources: {kernel_directory}/include/mos/dscosts.h. The other output file, mosix.costs, contains the 14 integer tuning parameters: when copied to the file /proc/mosix/admin/overheads, those parameters override the kernel's built-in parameters from dscosts.h. In paraticular, if mosix.costs is copied to the file /etc/overheads, the tuning parameters will be loaded at boot time (when MOSIX is being configured).

Similarly, mtune produces two output files, corresponding to two different methods of installing the parameters: The first output file, named mfscosts.h, is to be included as is in the kernel sources: {kernel_directory}/include/mos/mfscosts.h. The other output file, mfs.costs, contains the 6 integer tuning parameters: when copied to the file /proc/mosix/admin/mfscosts, those parameters override the kernel's built-in parameters from mfscosts.h. In paraticular, if mfs.costs is copied to the file /etc/mfscosts, the MFS tuning parameters will be loaded at boot time (when MOSIX is being configured).

tune_kernel is a more user-friendly front-end script to tune and mtune, guiding the system-administrator through the tuning procedure. There is no need for the common system-administrator to run tune (or mtune) directly (but if choosing to do so anyway, one must either run them in the /sbin directory, or first copy the file /sbin/tunepass to one's current directory). tune_kernel also supports complex network topologies, eg. when the network-distance is not even between all nodes, provided that the kernel is configured to support complex topologies. In that case, it combines the outputs of several invocations of tune and/or mtune to produce a complex version of /etc/overheads and/or /etc/mfscosts.

The pre-requisites for running tune and mtune successfully and obtaining accurate results are:

1.

Have two and only two nodes on the MOSIX network (either disconnect the others or turn them off). It is preferable, but not required that both nodes have the same hardware (otherwise, some manual intervention will be required by the program and the tuning process will take longer).

2

If any other MOSIX node(s) must remain connected to the network because they act as routers between the two tested nodes, then MOSIX must be disabled on those node(s) for the duration of the test (this can be done by running setpe -off).

3.

Place both nodes in single-user mode. One way to achieve this is by typing "telinit 1". The other way is to boot the nodes straight into single-user mode.

4

Run prep_tune on the other node (no need to run it on the node where you run tune_kerenl, since tune_kernel runs it there automatically). prep_tune does the following:

a.: enables the network, since it is normally disabled in single-user mode.
b.: enables/starts MOSIX.
c.: shuts down the (normally active) background operations of MOSIX, so that other ongoing system-activities do not interfere with the measurements.

4.

Do not touch either keyboard while tune is running.

USING THE OUTPUT OF TUNE(/MTUNE)

The output of tune (and mtune) is to be used only on the node where tune_kernel was running or on other nodes with identical hardware. If you have nodes of different hardware, you may need to run tune_kernel several times on different consoles.

In a simple cluster, where the network-distance is even between all nodes, both methods are available: either re-compiling the kernel with the produced dscosts.h (mfscosts.h for mtune) or copying the file mosix.costs, containing 14 tuning-parameters to /etc/overheads (or mfs.costs, containing 6 MFS tuning-parameters to /etc/mfscosts).

In a complex cluster, only the second method is available and MOSIX supports up to CONFIG_MOSIX_MAXTOPOLOGY (configurable in the kernel) different network distances from each node to CONFIG_MOSIX_MAXTOPOLOGY groups of other nodes. The /etc/overheads file may therefore contain up to (14+2)*CONFIG_MOSIX_MAXTOPOLOGY integer constants, where each 14 tuning-parameters are followed by 2 constants: FIRST and LAST.

Similarly, the /etc/mfscosts file may contain up to (6+2)*CONFIG_MOSIX_MAXTOPOLOGY integer constants, where each 6 MFS tuning-parameters are followed by 2 constants: FIRST and LAST.

*: When FIRST <= LAST the preceding 14 (or 6) parameters apply to all MOSIX-nodes numbered between FIRST and LAST (inclusive).
*: When FIRST > LAST the preceding 14 (or 6) parameters apply to all MOSIX-nodes NOT between LAST and FIRST.
*: When FIRST is non-zero and LAST is 0, the preceding 14 parameters (or 6) apply to all nodes except FIRST.
*: When FIRST is 0, the preceding 14 (or 6) parameters include all the remaining nodes.

The first set of 14 (or 6) parameters also applies to all other unselected nodes (if any).

FILES

/sbin/tunepass: A supplementary program required by tune.
/sbin/prep_tune: A script to prepare the system for tuning.
dscosts.h: An output of tune, to be included in the kernel.
mfscosts.h: An output of mtune, to be included in the kernel.
mosix.costs: An output of tune containing 14 integer tuning parameters.
mfs.costs: An output of mtune containing 6 integer MFS tuning parameters.
/proc/mosix/admin/overheads: Where the resulting tuning parameters are written (if different from the kernel's).
/proc/mosix/admin/mfscosts: Where the resulting MFS tuning parameters are written (if different from the kernel's).
/etc/overheads: Tuning parameters.
/etc/mfscosts: MFS Tuning parameters.

HISTORY

tune, mtune and tune_kernel are MOSIX utilities.