Using R on Netezza

R on Netezza's TwinFin Appliance
by Łukasz Bartnik, Netezza
lbartnik@netezza.com

Netezza & TwinFin
TwinFin, called also a “Data Warehouse Appliance” or “Analytic Appliance”, is a massively
parallel processing (MPP) database management system (DBMS). What distinguishes it among
other DBMSes, is the fact that it was designed from scratch, including the special hardware using
field programmable gate arrays (FPGAs), in order to attain the best performance and avoid the
pitfalls of parallel data processing.
Unlike TwinFin, other database platforms were designed initially to work on single computers – at
most multiprocessor, and bringing the parallel processing to them causes many problems, arising
from the different software architecture, etc. This is obviously the case always when a big and
complicated computer system is to be redesigned and adapted to a new hardware platform.
The main concepts behind TwinFin are:
• On-Stream AnalyticsTM – which can be expressed as “bring algorithms to data” instead of
bringing data to algorithms; this means that TwinFin comes with a set of mechanisms and
interfaces which support designing and writing parallel versions of algorithms utilizing the
specific capabilities of this platform
• no indices supporting SQL queries – all data processing is done on-the-fly and takes a huge
advantage of the FPGAs to filter out and extract only the desired data
• intelligent query optimizer – which prepares an execution plan with minimal effort put to
move data between the cluster nodes; the level of expertise required to implement such
optimizer is substantial, and this is one of the biggest advantages of the whole platform
Being a cluster, TwinFin consists of a number of nodes. One, called Host, is the master node user
connects to from his client computer; SQL interpreter & query optimizer work on the Host, also the
SQL stored procedures are executed on this node. Slave nodes, called SPUs (Snippet Processing
Units), are capable of executing SQL queries compiled on Host, UDXs (User Defined Functions
and Aggregates implemented in C++), they also are the actual place where the data is stored.

TwinFin & R
The TwinFin-R approach is quite straightforward. Since the data processing is handled in parallel, it
is obvious that also R has to comply with this paradigm. We have implemented a set of either
entirely new or new versions of previously existing mechanisms:
• apply & tapply – which are used to run a user-defined function on each row or each group of
rows given a grouping column, respectively
• nzrun, nzfilter, etc. – which are different, specialized (or generalized) versions of the two
above (more information below)
These functions are used with nz.data.frame – a Netezza's version of a well-known R data.frame
class. Basically, nz.data.frame is a pointer to a table which resides on the TwinFin. It implements a
number of methods for extracting a subset of its data, gathering meta-info, similar to the data.frame
ones, and working with parallel data-processing algorithms.
Sample code:

> library(nzr)
> nzconnect(“user”, “password”, “host”, “database”)
> library(rpart)
> data(kyphosis)
# this creates a table out of kyphosis data.frame
# and sends its data to TwinFin
> invisible(as.nz.data.frame(kyphosis))
> nzQuery("SELECT * FROM kyphosis")
KYPHOSIS AGE NUMBER START
1 absent 71 3 5
2 absent 158 3 14
3 present 128 4 5
[ cut ]
# now create a nz.data.frame
> k <- nz.data.frame(“kyphosis”)
> as.data.frame(k)
KYPHOSIS AGE NUMBER START
1 absent 71 3 5
2 absent 158 3 14
3 present 128 4 5
[ cut ]
> nzQuery(“SELECT * FROM kyphosis”)
COUNT
1 81

R on TwinFin – internals
The “R on TwinFin” package is built on top of RODBC: each of its functions/methods invoked in R
client is translated into SQL query and run directly on TwinFin. If a user-defined function has to be
applied on a TwinFin table, it is serialized and sent to TwinFin as a string literal.
There is a low-level API available on TwinFin, which is hidden by function like nzapply, nzfilter,
etc. When a user-defined code is run in parallel on TwinFin, on each node (SPU) a new R session is
started, which then proceeds with the code execution. Each such sessions has to communicate with
the DBMS process using the aforementioned low-level API to fetch & send data, report error, etc.
Below, are some of its functions:
• getNext() – fetches new row from the appliance
• getInputColumn(index) – return the value of the column specified by its index
• setOutput(index,value) – sets the output column value
• outputResult() – sends the current output row to the appliance
• columnCount() - returns the number of input columns
• userError() – reports an error
These functions might be used directly when nzrun is used to run a user-defined code/function, or
indirectly, through a high-level API, when nzapply or nztapply is used. The two following examples
show the same task approached in both ways.
> ndf <- nz.data.frame(“kyphosis”)
> fun <- function(x) {
> while(getNext()) {
> v <- inputColumn(0)
> setOutput(0, v ^ 2)
> outputResult()
> }
> }
> nzrun(ndf[,2], fun)
Example 1: low-level API, nzrun
> ndf <- nz.data.frame(“kyphosis”)
> nzapply(ndf[,2], NULL, function(x) x[[1]]^2)
Example 2: high-level API, nzapply

R on TwinFin – a bit more sophisticated
Given is a TwinFin table with transactions data; its columns are shop_id, prod_1, prod_2,
prod_3.The first column determined the shop where the transaction took place, the other three store
the total amount spent on three different product types. The goal is, for each shop, to build a linear
price model of the form:
prod_1 ~ prod_2 + prod_3
This can be accomplished with nztapply, where the grouping column will be shop_id:
> transactions <- nz.data.frame(“transactions”)
> prodlm <- function(x) {
> return(lm(prod_1 ~ prod_2 + prod_3, data=x)$coefficients)
> }
> nztapply(transactions, shop_id, prodlm)
This performs the linear-model-building procedure for each group determined by the shop_id
column. Groups are processed independently, and if they reside on different nodes (SPUs) due to
the data distribution specified for the TwinFin table, then the computations can take place in
parallel.
The transactions table can be generated with the function given in Appendix A.

Decision Trees – interfacing TwinFin algorithms in R
Since R implementation of computation-intensive algorithms might not be efficient enough, there is
a number of algorithms whose design and implementation (in SQL and C++) takes into
consideration properties specific to TwinFin.
One of them is the Decision Trees algorithm. It requires three datasets: training, testing, and
pruning. Then, a R wrapper dectree2 is called, which starts the computations on TwinFin. Object of
class dectree2 is returned as the result.
> adultTrain <- nz.data.frame("adult_train")
> adultTest <- nz.data.frame("adult_test")
> adultPrune <- nz.data.frame("adult_prune")
> tr <- dectree2(income~., data=adultTrain, valtable=adultPrune,
minsplit=1000, id="id", qmeasure="wAcc")
This can be then printed (Example 3) or plotted (see Fig. 1):
> print(tr)
node), split, n, deviance, yval, (yprob)
* denotes terminal node

16) root 0 0 small ( 0.5 0.5 ) *
8) education_num < 10 0 0 small ( 0.5 0.5 )
4) capital_gain < 5013 0 0 small ( 0.5 0.5 )
1) NANA 0 0 small ( 0.5 0.5 )
2) marital_status=Married-civ-spouse 0 0 small ( 0.5 0.5 )
17) education_num > 7 0 0 small ( 0.5 0.5 )
34) age < 35 0 0 small ( 0.5 0.5 ) *
140) hours_per_week < 34 0 0 small ( 0.5 0.5 ) *
35) age > 35 0 0 small ( 0.5 0.5 )
[ cut ]
Example 3: printing a TwinFin decision tree

Figure 1: Example plot of an dectree2 object

Appendix A
Generating data for the nztapply example in Section 4.
gendata <- function(prodno = 3, shopsno = 4, tno = 300) {
trans <- data.frame(matrix(NA, shopsno*tno, 4))
names(trans) <- c("shop_id", "prod_1", "prod_2", "prod_3")
r <- 1
for (s in seq(shopsno)) {
sigma <- matrix(NA, prodno, prodno)
for (t in seq(prodno)) {
x <- seq(prodno-t+1)+t-1
sigma[t,x] <- sigma[x,t] <- runif(length(x),max=0.75)
}
diag(sigma) <- 1
chsigma <- t(chol(sigma))
for (r in seq(tno)+tno*(s-1))
trans[r,] <- c(s,abs(chsigma %*% rnorm(prodno)))
}
return(trans)
}

Bibliography
1. http://www.netezza.com/data-warehouse-appliance-products/twinfin.aspx
2. http://www.netezza.com/documents/whitepapers/streaming_analytic_white_paper.pdf
3. http://www.netezza.com/analystReports/2006/bloor_100906.pdf

Using R on Netezza

Recommended

More Related Content

What's hot (20)

Viewers also liked (6)

Similar to Using R on Netezza (20)

More from Ajay Ohri (20)

Recently uploaded (20)

Using R on Netezza