Revision: @(#)ServConfig.txt  1.1     94/10/05

Server Configuration Requirements

This paper is based upon an article in Volume 5, Issue 3 of the SPEC 
Newsletter, by Brian Pawlowski, pp. 5-6.

The following paper expands on the run rule:

	For all architectures, SPEC intends that for every network, 
	all file systems should be accessed by all clients uniformly.  

to more fully explain the intention and motivation of the requirement, 
and to provide an example of how to produce conformant test setup.

The intent of this rule is to eliminate exploitation of any 
partitionable aspect of the benchmark in a test setup, particularly 
when reporting cluster results. When a vendor has determined their 
test setup topology for generating 097.LADDIS results, including the 
number of load generating processes desired, the assignment of load 
generators to file systems should be in a manner such that there is 
no inherent favoritism. SPEC recommends that a vendor make a best 
effort, for both clustered and non- clustered results reporting, to 
distribute file systems in such a way as to avoid having any load 
generator access only one processor, disk controller or disk (which 
makes up the file system). The following algorithm works well with 
the design of SFS 1.0:

	Once the number of load generating processes has been 
	determined, and their distribution to load generators 
	set (number of processes per load generator), then load 
	generator mount points distribute file systems in the 
	following manner: for each load generating process, 
	proceeding from load generator to load generator, and 
	network to  network should be assigned first to a 
	different cluster processor (if the configuration is a 
	cluster), then a different controller and then disk, 
	within the capabilities of the system; repeating 
	processor, controller or disk only after a load 
	generator has been assigned to each.

This rule reflects the consideration of clustered file servers, 
but can be simplified fur the non-clustered case. Rotating the file 
systems amongst the load generators, varying first the processor, 
then the controller and finally disk, eliminates any favoritism. 
SPEC recognize servers vary as to exposing elements such as 
processor, disk controller or disk to load generators remotely 
accessing file systems. The vendor needs to assure that a best 
effort has been made to avoid any preferred partitioning and file 
systems in the system under test. Other distribution schemes which 
avoid inherent favoritism are acceptable. 

The new run rule intends to prevent biased configurations for 
producing LADDIS results that interpreted several ways. Due to 
artifacts in the current implementation of LADDIS, certain 
interpretations are more problematical than others. A vendor 
should not have to modify the number of load generating processes 
they run to saturate a server, but should only modify how the 
distribute the file systems to the clients to avoid any inherent 
favoritism. While the thrust of the new run rule is intended for 
clustered reporting, it is to be applied to all reported results. 
For those cases where a sample rotation algorithm does not yield 
a uniform access configuration, the vendor must show intent to 
adhere to the new run rule. Some servers simply will not fit into 
the above discussion of distribution of file systems because of 
their design (which would not allow access to underlying details 
of controllers and disks - easily made unclear by using virtual 
file systems that span disks). In such cases, a vendor again must 
show intent.