In Bioconductor/HDF5Array: HDF5 datasets as array-like objects in R

Introduction

The aim of this document is to measure the performance of the r Biocpkg("HDF5Array") package for normalization and PCA (Principal Component Analysis) of on-disk single cell data, two computationally intensive operations at the core of single cell analysis.

The benchmarks presented in the document were specifically designed to observe the impact of two critical parameters on performance:

1. data representation (i.e. sparse vs dense);
2. size of the blocks used for block-processed operations.

Hopefully these benchmarks will also facilitate comparing performance of single cell analysis workflows based on r Biocpkg("HDF5Array") with workflows based on other tools like Seurat or Scanpy.

Install and load the required packages

Let's install and load r Biocpkg("HDF5Array") as well as the other packages used in this vignette:

if (!require("BiocManager", quietly=TRUE))
    install.packages("BiocManager")

pkgs <- c("HDF5Array", "ExperimentHub", "DelayedMatrixStats", "RSpectra")
BiocManager::install(pkgs)

Load the packages:

library(HDF5Array)
library(ExperimentHub)
library(DelayedMatrixStats)
library(RSpectra)

## Needed for the make_timings_table() function.
path <- system.file(package="HDF5Array",
                    "scripts", "make_timings_table.R", mustWork=TRUE)
source(path, verbose=FALSE)

The test datasets

Sparse vs dense representation

The datasets that we will use for our benchmarks are subsets of the 1.3 Million Brain Cell Dataset from 10x Genomics. This is a sparse 27,998 x 1,306,127 matrix of counts, with one gene per row and one cell per column. Around 7% of the matrix values are nonzero counts. The dataset is available via the r Biocpkg("ExperimentHub") package in two forms:

1. As a sparse HDF5 file: This is the original HDF5 file provided by 10x Genomics. It uses the CSR/CSC/Yale representation to store the sparse data.

2. As a dense HDF5 file: The same data as the above but stored as a regular HDF5 dataset with (compressed) chunks of dimensions 100 x 100.

The two files are hosted on r Biocpkg("ExperimentHub") under resource ids EH1039 and EH1040:

hub <- ExperimentHub()
hub["EH1039"]$description  # sparse representation
hub["EH1040"]$description  # dense representation

Let's download them to the local r Biocpkg("ExperimentHub") cache if they are not there yet:

## Note that this will be quick if the HDF5 files are already in the
## local ExperimentHub cache. Otherwise, it will take a while!
brain_s_path <- hub[["EH1039"]]
brain_D_path <- hub[["EH1040"]]

brain_s_path and brain_D_path are the paths to the downloaded files.

TENxMatrix vs HDF5Matrix objects

We use the TENxMatrix() and HDF5Array() constructors to bring the sparse and dense datasets in R, as DelayedArray derivatives. Note that this does not load the matrix data in memory.

Bring the sparse dataset in R

## Use 'h5ls(brain_s_path)' to find out the group.
brain_s <- TENxMatrix(brain_s_path, group="mm10")

brain_s is a 27,998 x 1,306,127 TENxMatrix object:

class(brain_s)
dim(brain_s)
is_sparse(brain_s)

See ?TENxMatrix in the r Biocpkg("HDF5Array") package for more information about TENxMatrix objects.

Bring the dense dataset in R

## Use 'h5ls(brain_D_path)' to find out the name of the dataset.
brain_D <- HDF5Array(brain_D_path, name="counts")

brain_D is a 27,998 x 1,306,127 HDF5Matrix object that contains the same data as brain_s:

class(brain_D)
dim(brain_D)
chunkdim(brain_D)
is_sparse(brain_D)

See ?HDF5Matrix in the r Biocpkg("HDF5Array") package for more information about HDF5Matrix objects.

Even though the data in brain_D_path is stored in a dense format, we can flag it as quantitatively sparse. This is done by calling the HDF5Array() constructor function with as.sparse=TRUE:

brain_Ds <- HDF5Array(brain_D_path, name="counts", as.sparse=TRUE)

The only difference between brain_D and brain_Ds is that the latter is now seen as a sparse object, and will be treated as such:

class(brain_Ds)
dim(brain_Ds)
chunkdim(brain_Ds)
is_sparse(brain_Ds)

Concretely this means that, when blocks of data are loaded from the dense HDF5 file to memory during block-processed operations, they end up directly in an in-memory sparse representation without going thru an in-memory dense representation first. This is expected to reduce memory footprint and (hopefully) will help with overall performance.

Finally note that the dense HDF5 file does not contain the dimnames of the matrix, so we manually add them to brain_s and brain_Ds:

dimnames(brain_Ds) <- dimnames(brain_D) <- dimnames(brain_s)

Create the test datasets

For our benchmarks below, we create subsets of the 1.3 Million Brain Cell Dataset of increasing sizes: subsets with 12,500 cells, 25,000 cells, 50,000 cells, 100,000 cells, and 200,000 cells. Note that subsetting a TENxMatrix or HDF5Matrix object with [ is a delayed operation so has virtually no cost:

brain1_s  <- brain_s[  , 1:12500]
brain1_D  <- brain_D[  , 1:12500]
brain1_Ds <- brain_Ds[ , 1:12500]

brain2_s  <- brain_s[  , 1:25000]
brain2_D  <- brain_D[  , 1:25000]
brain2_Ds <- brain_Ds[ , 1:25000]

brain3_s  <- brain_s[  , 1:50000]
brain3_D  <- brain_D[  , 1:50000]
brain3_Ds <- brain_Ds[ , 1:50000]

brain4_s  <- brain_s[  , 1:100000]
brain4_D  <- brain_D[  , 1:100000]
brain4_Ds <- brain_Ds[ , 1:100000]

brain5_s  <- brain_s[  , 1:200000]
brain5_D  <- brain_D[  , 1:200000]
brain5_Ds <- brain_Ds[ , 1:200000]

Block-processed normalization and PCA

Code used for normalization and PCA

We'll use the following code for normalization:

## Also selects the most variable genes (1000 by default).
simple_normalize <- function(mat, num_var_genes=1000)
{
    stopifnot(length(dim(mat)) == 2, !is.null(rownames(mat)))
    mat <- mat[rowSums(mat) > 0, ]
    col_sums <- colSums(mat) / 10000
    mat <- t(t(mat) / col_sums)
    row_vars <- rowVars(mat)
    row_vars_order <- order(row_vars, decreasing=TRUE)
    variable_idx <- head(row_vars_order, n=num_var_genes)
    mat <- log1p(mat[variable_idx, ])
    mat / rowSds(mat)
}

and the following code for PCA:

simple_PCA <- function(mat, k=25)
{
    stopifnot(length(dim(mat)) == 2)
    row_means <- rowMeans(mat)
    Ax <- function(x, args)
        (as.numeric(mat %*% x) - row_means * sum(x))
    Atx <- function(x, args)
        (as.numeric(x %*% mat) - as.vector(row_means %*% x))
    RSpectra::svds(Ax, Atrans=Atx, k=k, dim=dim(mat))
}

Block processing and block size

Note that the implementations of simple_normalize() and simple_PCA() are expected to work on any matrix-like object regardless of its exact type/representation e.g. it can be an ordinary matrix, a SparseMatrix object from the r Biocpkg("SparseArray") package, a dgCMatrix object from the r CRANpkg("Matrix") package, a DelayedMatrix derivative (TENxMatrix, HDF5Matrix, TileDBMatrix), etc...

However, when the input is a DelayedMatrix object or derivative, it's important to be aware that:

• Summarization methods like sum(), colSums(), rowVars(), or rowSds(), and matrix multiplication (%*%), are block-processed operations.

• The block size is 100 Mb by default. Increasing or decreasing the block size will typically increase or decrease the memory usage of block-processed operations. It will also impact performance, but sometimes in unexpected or counter-intuitive ways.

• The block size can be controlled with DelayedArray::getAutoBlockSize() and DelayedArray::setAutoBlockSize().

For our benchmarks below, we'll use the following block sizes:

| | NORMALIZATION | PCA | | -------------------- | ------------: | -----: | | TENxMatrix (sparse) | 250 Mb | 40 Mb | | HDF5Matrix (dense) | 16 Mb | 100 Mb | | HDF5Matrix as sparse | 250 Mb | 40 Mb |

Monitoring memory usage

While manually running our benchmarks below on a Linux or macOS system, we will also monitor memory usage at the command line in a terminal with:

(while true; do ps u -p <PID>; sleep 1; done) >ps.log 2>&1 &

where <PID> is the process id of our R session. This will allow us to measure the maximum amount of memory used by the calls to simple_normalize() or simple_PCA().

Normalization and PCA of the 27,998 x 12,500 test dataset

Normalization

In this section we run simple_normalize() on the three different representations (TENxMatrix, HDF5Matrix, and "HDF5Matrix as sparse") of the smaller test dataset only (27,998 x 12,500), and we report the time of each run.

See the Comprehensive timings obtained on various machines section below in this document for simple_normalize() and simple_pca() timings obtained on various machines on all our test datasets and using four different block sizes: 16 Mb, 40 Mb, 100 Mb, and 250 Mb.

TENxMatrix (sparse)

dim(brain1_s)
DelayedArray::setAutoBlockSize(250e6)  # set block size to 250 Mb
system.time(norm_brain1_s <- simple_normalize(brain1_s))
dim(norm_brain1_s)

HDF5Matrix (dense)

dim(brain1_D)
DelayedArray::setAutoBlockSize(16e6)   # set block size to 16 Mb
system.time(norm_brain1_D <- simple_normalize(brain1_D))
dim(norm_brain1_D)

HDF5Matrix as sparse

dim(brain1_Ds)
DelayedArray::setAutoBlockSize(250e6)  # set block size to 250 Mb
system.time(norm_brain1_Ds <- simple_normalize(brain1_Ds))
dim(norm_brain1_Ds)

On-disk realization of the normalized datasets

Note that the normalized datasets obtained in the previous section are DelayedMatrix objects that carry delayed operations. These operations can be displayed with showtree() e.g. for norm_brain1_s:

class(norm_brain1_s)
showtree(norm_brain1_s)

The other norm_brain1_* objects carry similar operations.

Before we proceed with PCA, we're going to write the normalized datasets to new HDF5 files. This introduces an additional cost, but, on the other hand, it has the benefit of triggering on-disk realization of the object. This means that all the delayed operations carried by the object will get realized on-the-fly before the matrix data actually lands on the disk, making the new object (TENxMatrix or HDF5Matrix) more efficient for PCA or whatever block-processed operations will come next.

We will use blocks of 100 Mb for all the writing operations.

DelayedArray::setAutoBlockSize(1e8)

TENxMatrix (sparse)

dim(norm_brain1_s)
system.time(norm_brain1_s <- writeTENxMatrix(norm_brain1_s, level=0))

The new norm_brain1_s object is a pristine TENxMatrix object:

class(norm_brain1_s)
showtree(norm_brain1_s)  # "pristine" object (i.e. no more delayed operations)

HDF5Matrix (dense)

dim(norm_brain1_D)
system.time(norm_brain1_D <- writeHDF5Array(norm_brain1_D, level=0))

The new norm_brain1_D object is a pristine HDF5Matrix object:

class(norm_brain1_D)
showtree(norm_brain1_D)  # "pristine" object (i.e. no more delayed operations)

HDF5Matrix as sparse

dim(norm_brain1_Ds)
system.time(norm_brain1_Ds <- writeHDF5Array(norm_brain1_Ds, level=0))

The new norm_brain1_Ds object is a pristine sparse HDF5Matrix object:

class(norm_brain1_Ds)
showtree(norm_brain1_Ds)  # "pristine" object (i.e. no more delayed operations)

PCA

In this section we run simple_pca() on the normalized datasets obtained in the previous section and report the time of each run.

See the Comprehensive timings obtained on various machines section below in this document for simple_normalize() and simple_pca() timings obtained on various machines on all our test datasets and using four different block sizes: 16 Mb, 40 Mb, 100 Mb, and 250 Mb.

TENxMatrix (sparse)

DelayedArray::setAutoBlockSize(40e6)   # set block size to 40 Mb
dim(norm_brain1_s)
system.time(pca1s <- simple_PCA(norm_brain1_s))

HDF5Matrix (dense)

DelayedArray::setAutoBlockSize(1e8)    # set block size to 100 Mb
dim(norm_brain1_D)
system.time(pca1D <- simple_PCA(norm_brain1_D))

Sanity check:

stopifnot(all.equal(pca1D, pca1s))

HDF5Matrix as sparse

DelayedArray::setAutoBlockSize(40e6)   # set block size to 40 Mb
dim(norm_brain1_Ds)
system.time(pca1Ds <- simple_PCA(norm_brain1_Ds))

Sanity check:

stopifnot(all.equal(pca1Ds, pca1s))

Comprehensive timings obtained on various machines

Here we report timings (and memory usage) observed on various machines. For each machine, the results are presented in a table that shows the normalization & realization & PCA timings obtained for all our test datasets and using four different block sizes: 16 Mb, 40 Mb, 100 Mb, and 250 Mb. For each operation, the best time across the four different block sizes is displayed in bold.

All the timings (and memory usage) were produced by running the run_benchmarks.sh script located in the HDF5Array/inst/scripts/ folder of the package, using R 4.5 and r Biocpkg("HDF5Array") 1.35.12 (Bioconductor 3.21).

Timings for DELL XPS 15 laptop

hdparm1 <- "Output of <code>sudo hdparm -Tt &lt;device&gt;</code>:"
hdparm1 <- sprintf("<span style=\"font-style: italic\">%s</span>", hdparm1)
hdparm2 <- c(
    "Timing cached reads:   35188 MB in  2.00 seconds = 17620.75 MB/sec",
    "Timing buffered disk reads: 7842 MB in  3.00 seconds = 2613.57 MB/sec"
)
hdparm2 <- sprintf("<code>%s</code>", paste(hdparm2, collapse="<br />"))
disk_perf <- paste0(hdparm1, "<br />", hdparm2)
make_machine_specs_table("Specs for DELL XPS 15 laptop (model 9520)",
    specs=c(`OS`="Linux Ubuntu 24.04",
            `RAM`="32GB",
            `Disk`="1TB SSD"),
    disk_perf=disk_perf)

caption <- "Table 1: Timings for DELL XPS 15 laptop"
make_timings_table("xps15", caption=caption)

Timings for Supermicro SuperServer 1029GQ-TRT

hdparm1 <- "Output of <code>sudo hdparm -Tt &lt;device&gt;</code>:"
hdparm1 <- sprintf("<span style=\"font-style: italic\">%s</span>", hdparm1)
hdparm2 <- c(
    "Timing cached reads:   20592 MB in  1.99 seconds = 10361.41 MB/sec",
    "Timing buffered disk reads: 1440 MB in  3.00 seconds = 479.66 MB/sec"
)
hdparm2 <- sprintf("<code>%s</code>", paste(hdparm2, collapse="<br />"))
disk_perf <- paste0(hdparm1, "<br />", hdparm2)
make_machine_specs_table("Specs for Supermicro SuperServer 1029GQ-TRT",
    specs=c(`OS`="Linux Ubuntu 22.04",
            `RAM`="384GB",
            `Disk`="1.3TB ATA Disk"),
    disk_perf=disk_perf)

caption <- "Table 2: Timings for Supermicro SuperServer 1029GQ-TRT"
make_timings_table("rex3", caption=caption)

Timings for Apple Silicon Mac Pro

make_machine_specs_table("Specs for Apple Silicon Mac Pro (Apple M2 Ultra)",
    specs=c(`OS`="macOS 13.7.1",
            `RAM`="192GB",
            `Disk`="2TB SSD"),
    disk_perf="N/A")

caption <- "Table 3: Timings for Apple Silicon Mac Pro"
make_timings_table("kjohnson3", caption=caption)

Timings for Intel Mac Pro

make_machine_specs_table("Specs for Intel Mac Pro (24-Core Intel Xeon W)",
    specs=c(`OS`="macOS 12.7.6",
            `RAM`="96GB",
            `Disk`="1TB SSD"),
    disk_perf="N/A")

caption <- "Table 4: Timings for Intel Mac Pro"
make_timings_table("lconway", caption=caption)

Discussion

The "[Ds] HDF5Matrix as sparse" representation didn't live up to its promise so we leave it alone for now. Note that the code for loading blocks of data from the dense HDF5 file to memory does not currently take full advantage of the SVT_SparseArray representation, a new efficient data structure for multidimensional sparse data implemented in the r Biocpkg("SparseArray") package that overcomes some of the limitations of the dgCMatrix representation from the r CRANpkg("Matrix") package. This will need to be addressed.

TENxMatrix (sparse) vs HDF5Matrix (dense)

Normalization

There's no obvious best choice between the "[s] TENxMatrix (sparse)" and "[D] HDF5Matrix (dense)" representations. More precisely, for normalization, the former tends to give the best times when using bigger blocks (e.g. 250 Mb), whereas the latter tends to give the best times when using smaller blocks (e.g. 16 Mb).

Therefore, on a machine with enough memory to support big block sizes, one will get the best results with the [s] representation and blocks of 250 Mb or more. However, on a machine with not enough memory to support such big blocks, one should instead use the [D] representation with blocks of 16 Mb.

[TODO: Add some plots to help vizualize the above observations.]

PCA

For PCA, choosing the "[s] TENxMatrix (sparse)" representation and using small block sizes (40 Mb) tends to give the best performance.

[TODO: Add some plots to help vizualize this observation.]

Hybrid approach

Note that, on a machine where using blocks of 250 Mb or more for normalization is not an option, one should use the following hybrid approach:

• Start with the "[D] HDF5Matrix (dense)" representation.

• Perform normalization, using very small blocks (16 Mb).

• Switch to the "[s] TENxMatrix (sparse)" format when writing the normalized dataset to disk, that is, use writeTENxMatrix() instead of writeHDF5Array() for on-disk realization of the intermediate dataset.

• Perform PCA on the object returned by the on-disk realization step (writeTENxMatrix()), using small blocks (40 Mb).

A note about memory usage

The machines running macOS use between 2x and 3x more memory than the machines running Linux for the same task using the same block size.

Overall, on Linux, and for a given choice of block size, memory usage doesn't seem to be affected too much by the number of cells in the dataset, that is, all operations seem to perform at almost constant memory.

However, the "[D] HDF5Matrix (dense)" representation appears to be better than the "[s] TENxMatrix (sparse)" representation at keeping memory usage (mostly) flat as the number of cells in the dataset increases. This is even more accentuated on macOS where, somehow counter intuitively, using the dense representation manages to keep memory usage at a reasonable level (and more or less capped with respect to the number of cells), while using the sparse representation fails to do that.

Main takeaways

1. Using the TENxMatrix representation (sparse format), one can perform normalization and PCA of the bigger dataset (200,000 cells) on an average consumer-grade laptop like the DELL XPS 15 laptop (model 9520) in less than 25 minutes and using less than 4Gb of memory, as shown in the table below. For comparison, normalization and PCA on an in-memory representation of the same dataset (e.g. on a SparseMatrix object) takes less then 4 minutes. However, that consumes about 18.5Gb of memory! This will be the subject of an upcoming vignette in the r Biocpkg("SparseArray") package.

machine_names <- c(
    `DELL XPS 15 laptop`="xps15",
    `Supermicro SuperServer 1029GQ-TRT`="rex3",
    `Apple Silicon Mac Pro`="kjohnson3",
    `Intel Mac Pro`="lconway"
)
summarize_machine_times(machine_names)

1. Normalization and PCA are roughly linear in time with respect to the number of cells in the dataset, regardless of representation (sparse or dense) or block size.

2. Block size matters. When using the TENxMatrix representation (sparse format), the bigger the blocks the faster normalization will be (at the cost of increased memory usage). On the other hand, it seems that PCA prefers small blocks, at least with our naive simple_PCA() implementation.

3. Disk performance is important (not surprisingly) as attested by the lower performance of the Supermicro SuperServer 1029GQ-TRT machine, likely due to its slower disk.

Session information

sessionInfo()

Bioconductor/HDF5Array documentation built on June 8, 2025, 4:19 a.m.