从列表到data.table与hash的R快速单项查找

Question

One of the problems I often face is needing to look up an arbitrary row from a data.table. I ran into a problem yesterday where I was trying to speed up a loop and using profvis I found that the lookup from the data.table was the most costly part of the loop. I then decided to try and find the fastest way to do a single item lookup in R.

The data often takes the form of a data.table with a key column of the character type. The remaining columns are typically numeric values. I tried to create a random table with similar characteristics to what I often deal with which means >100K rows. I compared the native list, data.table package and the hash package. The native list and data.table were comparable for individual item lookup performance. Hash appeared to be two orders of magnitude faster. The tests were made up of 10 sets of 10,000 keys randomly sampled to provide for variance in access behavior. Each lookup method used the same sets of keys.

Ultimately my preference would be to either get the row lookup for data.table to be faster instead of having to create a hash table of my data or establish that it cannot be done and just use the hash package when I have to do fast lookup. I don't know if it would be possible but could you create a hash table of references to the rows in the data.table to allow for fast lookup using the hash package? I know that type of thing is possible in C++ but to my knowledge R does not allow this kind of thing due to the lack of pointers.

To Summarize: 1) Am I using data.table correctly for the lookups and therefore this is the speed I should expect for a single row lookup? 2) Would it be possible to create a hash of pointers to the data.table rows to allow for fast lookup that way?

Test System:

Windows 10 Pro x64

R 3.2.2

data.table 1.9.6

hash 2.2.6

Intel Core i7-5600U with 16 GB RAM

Code:

library(microbenchmarkCore) # install.packages("microbenchmarkCore", repos="http://olafmersmann.github.io/drat")
library(data.table)
library(hash)

# Set seed to 42 to ensure repeatability
set.seed(42)

# Setting up test ------

# Generate product ids
product_ids <- as.vector(
  outer(LETTERS[seq(1, 26, 1)],
    outer(outer(LETTERS[seq(1, 26, 1)], LETTERS[seq(1, 26, 1)], paste, sep=""),
          LETTERS[seq(1, 26, 1)], paste, sep = ""
    ), paste, sep = ""
  )
)

# Create test lookup data
test_lookup_list <- lapply(product_ids, function(id){
  return_list <- list(
    product_id = id,
    val_1 = rnorm(1),
    val_2 = rnorm(1),
    val_3 = rnorm(1),
    val_4 = rnorm(1),
    val_5 = rnorm(1),
    val_6 = rnorm(1),
    val_7 = rnorm(1),
    val_8 = rnorm(1)
  )
  return(return_list)
})

# Set names of items in list
names(test_lookup_list) <- sapply(test_lookup_list, function(elem) elem[['product_id']])

# Create lookup hash
lookup_hash <- hash(names(test_lookup_list), test_lookup_list)

# Create data.table from list and set key of data.table to product_id field
test_lookup_dt <- rbindlist(test_lookup_list)
setkey(test_lookup_dt, product_id)

test_lookup_env <- list2env(test_lookup_list)

# Generate sample of keys to be used for speed testing
lookup_tests <- lapply(1:10, function(x){
  lookups <- sample(test_lookup_dt$product_id, 10000)
  return(lookups)
})

# Native list timing
native_list_timings <- sapply(lookup_tests, function(lookups){
  timing <- system.nanotime(
    for(lookup in lookups){
      return_value <- test_lookup_list[[lookup]]
    }    
  )
  return(timing[['elapsed']])
})

# Data.table timing
datatable_timings <- sapply(lookup_tests, function(lookups){
  timing <- system.nanotime(
    for(lookup in lookups){
      return_value <- test_lookup_dt[lookup]
    }
  )
  return(timing[['elapsed']])
})


# Hashtable timing
hashtable_timings <- sapply(lookup_tests, function(lookups){
  timing <- system.nanotime(
    for(lookup in lookups){
      return_value <- lookup_hash[[lookup]]
    }
  )
  return(timing[['elapsed']])
})

# Environment timing
environment_timings <- sapply(lookup_tests, function(lookups){
  timing <- system.nanotime(
    for(lookup in lookups){
      return_value <- test_lookup_env[[lookup]]
    }
  )
  return(timing[['elapsed']])
})

# Summary of timing results
summary(native_list_timings)
summary(datatable_timings)
summary(hashtable_timings)
summary(environment_timings)

These were the results:

> # Summary of timing results
> summary(native_list_timings)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
  35.12   36.20   37.28   37.05   37.71   39.24 
> summary(datatable_timings)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
  49.13   51.51   52.64   52.76   54.39   55.13 
> summary(hashtable_timings)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
 0.1588  0.1857  0.2107  0.2213  0.2409  0.3258 
> summary(environment_timings)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
0.09322 0.09524 0.10680 0.11850 0.13760 0.17140 

It appears that the hash lookup is approximately two orders of magnitude faster than either the native list or data.table in this particular scenario.

Update: 2015-12-11 3:00PM PST

I received feedback from Neal Fultz suggesting the use of the native Environment object. Here is the code and result I got:

test_lookup_env <- list2env(test_lookup_list)
# Environment timing
environment_timings <- sapply(lookup_tests, function(lookups){
  timing <- system.nanotime(
    for(lookup in lookups){
      return_value <- test_lookup_env[[lookup]]
    }
  )
  return(timing[['elapsed']])
})
summary(environment_timings)
> summary(environment_timings)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
0.09322 0.09524 0.10680 0.11850 0.13760 0.17140 

Indeed, it appears that Environment is faster for individual item access in this scenario. Thank you Neal Fultz for pointing this method out. I appreciate having a more thorough understanding of the object types available in R. My questions still stand: am I using data.table correctly (I expect so but I am open to critique) and is there a way to provide row access to the rows of a data.table using some kind of pointer magic which would provide faster individual row access.

Clarification: 2015-12-11 3:52 PM PST

There have been some mentions that my access pattern in the inner-most loop of my test is inefficient. I agree. What I am trying to do is emulate as closely as possible the situation that I am dealing with. The loop that this is actually occurring in does not allow for vectorization which is why I am not using it. I realize this is not strictly the 'R' way of doing things. The data.table in my code is providing reference information and I do not necessarily know which row I need until I am inside the loop which is why I am trying to figure out how to access an individual item as quickly as possible, preferably with the data still stored in a data.table . This is also in part a curiosity question, can it be done?

Update 2: 2015-12-11 4:12 PM PST

I received feedback from @jangrorecki that using Sys.time() is an ineffective means of measuring the performance of a function. I have since revised the code to use system.nanotime() per the suggestion. The original code has been updated and the timing results.

The question still stands: is this the fastest way to do a row lookup of a data.table and if so is it possible to create a hash of pointers to the rows for quick lookup? At this point I am most curious how far R can be pushed. As someone who came from C++, this is a fun challenge.

Conclusion

I accepted the answer provided by Neal Fultz because it discussed what I was actually wanting to know. That said, this is not the way data.table was intended to be used so no one should interpret this to mean data.table is slow, it is actually incredibly fast. This was a very particular use case that I was curious about. My data comes in as a data.table and I wanted to know if I could get quick row access while leaving it as a data.table . I also wanted to compare the data.table access speed with a hash-table which is what is often used for fast, non-vectorized item lookup.

Answer 1

For a non-vectorized access pattern, you might want to try the builtin environment objects:

require(microbenchmark)

test_lookup_env <- list2env(test_lookup_list)


x <- lookup_tests[[1]][1]
microbenchmark(
    lookup_hash[[x]],
    test_lookup_list[[x]],
    test_lookup_dt[x],
    test_lookup_env[[x]]
)

Here you can see it's even zippier than hash :

Unit: microseconds
                  expr      min        lq       mean    median        uq      max neval
      lookup_hash[[x]]   10.767   12.9070   22.67245   23.2915   26.1710   68.654   100
 test_lookup_list[[x]]  847.700  853.2545  887.55680  863.0060  893.8925 1369.395   100
     test_lookup_dt[x] 2652.023 2711.9405 2771.06400 2758.8310 2803.9945 3373.273   100
  test_lookup_env[[x]]    1.588    1.9450    4.61595    2.5255    6.6430   27.977   100

EDIT:

Stepping through data.table:::`[.data.table` is instructive why you are seeing dt slow down. When you index with a character and there is a key set, it does quite a bit of bookkeeping, then drops down into bmerge , which is a binary search. Binary search is O(log n) and will get slower as n increases.

Environments, on the other hand, use hashing (by default) and have constant access time with respect to n.

To work around, you can manually build a map and index through it:

x <- lookup_tests[[2]][2]

e <- list2env(setNames(as.list(1:nrow(test_lookup_dt)), test_lookup_dt$product_id))

#example access:
test_lookup_dt[e[[x]], ]

However, seeing so much bookkeeping code in the data.table method, I'd try out plain old data.frames as well:

test_lookup_df <- as.data.frame(test_lookup_dt)

rownames(test_lookup_df) <- test_lookup_df$product_id

If we are really paranoid, we could skip the [ methods altogether and lapply over the columns directly.

Here are some more timings (from a different machine than above):

> microbenchmark(
+   test_lookup_dt[x,],
+   test_lookup_dt[x],
+   test_lookup_dt[e[[x]],],
+   test_lookup_df[x,],
+   test_lookup_df[e[[x]],],
+   lapply(test_lookup_df, `[`, e[[x]]),
+   lapply(test_lookup_dt, `[`, e[[x]]),
+   lookup_hash[[x]]
+ )
Unit: microseconds
                                expr       min         lq        mean     median         uq       max neval
                 test_lookup_dt[x, ]  1658.585  1688.9495  1992.57340  1758.4085  2466.7120  2895.592   100
                   test_lookup_dt[x]  1652.181  1695.1660  2019.12934  1764.8710  2487.9910  2934.832   100
            test_lookup_dt[e[[x]], ]  1040.869  1123.0320  1356.49050  1280.6670  1390.1075  2247.503   100
                 test_lookup_df[x, ] 17355.734 17538.6355 18325.74549 17676.3340 17987.6635 41450.080   100
            test_lookup_df[e[[x]], ]   128.749   151.0940   190.74834   174.1320   218.6080   366.122   100
 lapply(test_lookup_df, `[`, e[[x]])    18.913    25.0925    44.53464    35.2175    53.6835   146.944   100
 lapply(test_lookup_dt, `[`, e[[x]])    37.483    50.4990    94.87546    81.2200   124.1325   241.637   100
                    lookup_hash[[x]]     6.534    15.3085    39.88912    49.8245    55.5680   145.552   100

Overall, to answer your questions, you are not using data.table "wrong" but you are also not using it in the way it was intended (vectorized access). However, you can manually build a map to index through and get most of the performance back.

Answer 2

The approach you have taken seems to be very inefficient because you are querying multiple times the single value from the dataset.

It would be much more efficient to query all of them at once and then just loop on the whole batch, instead of querying 1e4 one by one.

See dt2 for a vectorized approach. Still it is hard for me to imagine the use case for that.

Another thing is 450K rows of data is quite few to make a reasonable benchmark, you may get totally different results for 4M or higher. In terms of hash approach you would probably also hit memory limits faster.

Additionally the Sys.time() may not be the best way to measure timing, read gc argument in ?system.time .

Here is the benchmark I've made using the system.nanotime() function from microbenchmarkCore package.

It is possible to speed up data.table approach even further by collapsing test_lookup_list into data.table and performing merge to test_lookup_dt , but to compare to hash solution I would also need to preprocess it.

library(microbenchmarkCore) # install.packages("microbenchmarkCore", repos="http://olafmersmann.github.io/drat")
library(data.table)
library(hash)

# Set seed to 42 to ensure repeatability
set.seed(42)

# Setting up test ------

# Generate product ids
product_ids = as.vector(
    outer(LETTERS[seq(1, 26, 1)],
          outer(outer(LETTERS[seq(1, 26, 1)], LETTERS[seq(1, 26, 1)], paste, sep=""),
                LETTERS[seq(1, 26, 1)], paste, sep = ""
          ), paste, sep = ""
    )
)

# Create test lookup data
test_lookup_list = lapply(product_ids, function(id) list(
    product_id = id,
    val_1 = rnorm(1),
    val_2 = rnorm(1),
    val_3 = rnorm(1),
    val_4 = rnorm(1),
    val_5 = rnorm(1),
    val_6 = rnorm(1),
    val_7 = rnorm(1),
    val_8 = rnorm(1)
))

# Set names of items in list
names(test_lookup_list) = sapply(test_lookup_list, `[[`, "product_id")

# Create lookup hash
lookup_hash = hash(names(test_lookup_list), test_lookup_list)

# Create data.table from list and set key of data.table to product_id field
test_lookup_dt <- rbindlist(test_lookup_list)
setkey(test_lookup_dt, product_id)

# Generate sample of keys to be used for speed testing
lookup_tests = lapply(1:10, function(x) sample(test_lookup_dt$product_id, 1e4))

native = lapply(lookup_tests, function(lookups) system.nanotime(for(lookup in lookups) test_lookup_list[[lookup]]))
dt1 = lapply(lookup_tests, function(lookups) system.nanotime(for(lookup in lookups) test_lookup_dt[lookup]))
hash = lapply(lookup_tests, function(lookups) system.nanotime(for(lookup in lookups) lookup_hash[[lookup]]))
dt2 = lapply(lookup_tests, function(lookups) system.nanotime(test_lookup_dt[lookups][, .SD, 1:length(product_id)]))

summary(sapply(native, `[[`, 3L))
#   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
#  27.65   28.15   28.47   28.97   28.78   33.45
summary(sapply(dt1, `[[`, 3L))
#   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
#  15.30   15.73   15.96   15.96   16.29   16.52
summary(sapply(hash, `[[`, 3L))
#   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
# 0.1209  0.1216  0.1221  0.1240  0.1225  0.1426 
summary(sapply(dt2, `[[`, 3L))
#   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
#0.02421 0.02438 0.02445 0.02476 0.02456 0.02779