anonymous_namespace{cbr.cpp} Namespace Reference

Classes
struct	Read
struct	Write
struct	Time
struct	RO
struct	RW
struct	R_RO
struct	R_Time
struct	W_RO
struct	W_Time
struct	Time_RO
struct	R_W_RO
struct	R_W_Time
struct	R_Time_RO
struct	W_Time_RO
struct	R_W_Time_RO
struct	TxnRatio
struct	TxnRatio_R
struct	TxnRatio_W
struct	TxnRatio_RO
struct	TxnRatio_Time
struct	TxnRatio_RW
struct	TxnRatio_R_RO
struct	TxnRatio_R_Time
struct	TxnRatio_W_RO
struct	TxnRatio_W_Time
struct	TxnRatio_RO_Time
struct	TxnRatio_RW_RO
struct	TxnRatio_RW_Time
struct	TxnRatio_R_RO_Time
struct	TxnRatio_W_RO_Time
struct	TxnRatio_RW_RO_Time

Functions
TM_INLINE unsigned long	norm_dist (unsigned long a, unsigned long b)
TM_FASTCALL uint32_t	profile_nochange ()
template<class C >
uint32_t	cbr_tail (qtable_t &profile)
TM_FASTCALL uint32_t	pol_CBR_RO ()
template<class C >
TM_FASTCALL uint32_t	cbr_nn ()

Detailed Description

This file implements a few classifiers for trying to pick which algorithm to switch to, based on some existing information from a profile and a table of previous experiments.

Function Documentation

template<class C >

TM_FASTCALL uint32_t anonymous_namespace{cbr.cpp}::cbr_nn

(

)

This template provides the functionality we need for doing a nearest-neighbor classification where we use as input the summary profile, and scan through the qtable to find the most representative row

Here is the call graph for this function:

template<class C >

uint32_t anonymous_namespace{cbr.cpp}::cbr_tail

(

qtable_t &

profile

)

Common part of CBR code

TM_INLINE unsigned long anonymous_namespace{cbr.cpp}::norm_dist	(	unsigned long	a,
		unsigned long	b
	)

The CBR policies can all be characterized as follows:

• We have a qtable full of normalized profiles
• We have a summary_profile, which is the average of N profiles, normalized
• For each qtable row, for each column, compute the norm_dist between the row and the summary_profile. Take the weighted sum of the distances

• Return the algorithm whose name appears in the qtable row with the minimum weighted sum norm_dist

  Implicit in this description is the fact that norm_dist must be defined relative to a normalization strategy. Using this norm_dist function implies that we tend to choose a qtable entry with larger number on some attribute when we have a tie in Manhattan distance. E.g., when we get a profiled read number of 4, we may only have 2 entries in the qtable, one with a read number of 3 and the other one with a read number of 5. In this case, we will choose the row with 5.

Here is the caller graph for this function:

TM_FASTCALL uint32_t anonymous_namespace{cbr.cpp}::pol_CBR_RO

(

)

This policy compares the read-only ratio of the current workload to the read-only ratios of entries in our qtable, using a nearest neighbor distance. Note that it does not use profiles at all.

Here is the caller graph for this function:

TM_FASTCALL uint32_t anonymous_namespace{cbr.cpp}::profile_nochange

(

)

A testing policy: we use this to return exactly the same algorithm as we had before a profile was run. The reason this is implemented here is just so that we have a hook in the standard place to ensure that profiles get printed.

Here is the call graph for this function:

Here is the caller graph for this function: