Skip to content

Commit

Permalink
Update paper
Browse files Browse the repository at this point in the history
  • Loading branch information
@ValHayot
ValHayot committed Mar 13, 2024
1 parent 8ab47dc commit 552b63b
Show file tree
Hide file tree
Showing 2 changed files with 84 additions and 12 deletions.
20 changes: 8 additions & 12 deletions paper/sea-neuro/paper-neuro.tex
View file
Original file line number Diff line number Diff line change
Expand Up @@ -205,9 +205,9 @@
\end{figure*}
Sea (Figure~\ref{fig:seaneuro:diagram}) is a data-management library that leverages
the \texttt{LD\_PRELOAD}
trick to intercept POSIX file system calls (more specifically, file access calls to the GNU C library,
glibc) on Linux systems. This enables Sea to redirect write calls
aimed at slower storage devices to a faster devices whenever possible.
trick to intercept POSIX file system calls, or more specifically, file access calls to the GNU C library,
glibc, on Linux systems. This enables Sea to redirect write calls
aimed at slower storage devices to a faster device whenever possible.
Similarly, when intercepting read calls, Sea can choose to read from a
faster device if a copy is available on that device. Sea decides which
storage location it can write to based on the details provided in an
Expand Down Expand Up @@ -300,11 +300,9 @@ \subsection{Speedups observed in controlled environment}
performance had been degraded by busy writers (Figure~\ref{fig:seaneuro:slashbin}),
although performance was variable. On average, a single SPM pipeline
process preprocessing a single HCP image produced the greatest speedup
(13$\times$) \TG{I think this number should be 32x, according to the current notebook}. The AFNI pipeline preprocessing a single fMRI image of the PREVENT-AD dataset
(12.6$\times$), \TG{I think this number should be 32x, according to the current notebook}. The AFNI pipeline preprocessing a single fMRI image of the PREVENT-AD dataset
using a single process was the next fastest pipeline, with an average speedup of
(5$\times$). The FSL Feat pipeline had speedups as well, with a maximum average speedup of
1.3$\times$ when preprocessing a single PREVENT-AD image. The SPM pipeline consistently had
(5$\times$). The FSL Feat pipeline had speedups as well, with a maximum average speedup of
(4.3$\times$). The FSL Feat pipeline had speedups as well, with a maximum average speedup of
1.3$\times$ when preprocessing a single PREVENT-AD image. The SPM pipeline consistently had
excellent speedups, which is likely due to a mix of prefetching the initial
input files and the I/O patterns of the application. While it was expected that the
Expand All @@ -322,13 +320,12 @@ \subsection{Speedups observed in controlled environment}
The FSL Feat pipeline, in contrast, appeared to be the most compute-bound of the three
applications. Not only did it spend an extensive amount of time computing, the
amount of output data generated was least of the three. Due to the
compute-intensive nature of the FSL Feat pipeline, it is expected that it resulted in the
compute-intensive nature of the FSL Feat pipeline, it is expected that it resulted in the
compute-intensive nature of the FSL Feat pipeline, it is expected to have resulted in the
least significant speedups.

The HCP dataset, on average, obtained the greatest speedups with
Sea \TG{Mention value}. Out of the three datasets, HCP has the largest images (see
Table~\ref{table:sea-neuro:data}). Larger individual images mean that they
Table~\ref{table:sea-neuro:data}). Larger individual images
occupy more page cache space and take longer to flush to Lustre. Unsurprisingly,
the next dataset with the largest images (\SI{282}{\mebi\byte} compressed for a
single image) results in the next largest speedups, followed by PREVENT-AD, the
Expand All @@ -349,8 +346,7 @@ \subsection{Speedups observed in controlled environment}
\subsection{Speedup correlated with Lustre degradation}

Baseline performance without busy writers was comparable to that of Sea's as can be seen in Figure~\ref{fig:seaneuro:slashbin} \TG{(p=x, two-sample T-test)}. However, with
busy writers, the makespan obtained with Sea was smaller than baseline \TG{(p=x, two-sample T-test)}. We observe that the more Baseline deviates from this identity line, the
greater the speedup \TG{I would remove this sentence, it's obvious}. We did occasionally observe slowdowns from using Sea. These slowdowns
busy writers, the makespan obtained with Sea was smaller than baseline \TG{(p=x, two-sample T-test)}. We did occasionally observe slowdowns from using Sea. These slowdowns
may arise from the initial read of the data, as they appeared to occur less
frequently with the SPM pipeline, or due to increased CPU contention caused by Sea's rapid I/O.

Expand Down
76 changes: 76 additions & 0 deletions paper/sea-neuro/results/figure_slashbin.ipynb
View file
Original file line number Diff line number Diff line change
Expand Up @@ -6898,6 +6898,82 @@
"df_merged.sort_values(by=\"speed_up\", ascending=False)"
]
},
{
"cell_type": "code",
"execution_count": 47,
"id": "0f35c653",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"dataset pipeline n_images busy_writers\n",
"HCP spm 1 6 12.617431\n",
"ds001545 spm 8 6 7.615075\n",
" 1 6 6.571933\n",
"HCP spm 8 6 5.160400\n",
"preventAD afni 1 6 4.271394\n",
"HCP afni 8 6 4.074774\n",
"ds001545 spm 16 6 3.585750\n",
"HCP spm 16 6 2.937354\n",
"preventAD spm 1 6 2.828866\n",
"ds001545 afni 1 6 2.369637\n",
"HCP afni 16 6 2.291281\n",
"preventAD afni 8 6 1.622492\n",
"HCP afni 1 6 1.475868\n",
"preventAD fsl 1 6 1.345780\n",
"ds001545 fsl 1 6 1.261522\n",
" afni 16 6 1.150132\n",
"preventAD fsl 8 6 1.121890\n",
" afni 16 0 1.079569\n",
"ds001545 spm 1 0 1.064818\n",
"preventAD spm 16 6 1.059949\n",
"HCP spm 1 0 1.057301\n",
" afni 8 0 1.053507\n",
"ds001545 fsl 8 6 1.039927\n",
"preventAD afni 1 0 1.034140\n",
"HCP afni 1 0 1.033463\n",
"preventAD spm 1 0 1.033344\n",
" fsl 16 6 1.031404\n",
" spm 8 0 1.024793\n",
"ds001545 afni 1 0 1.022089\n",
" spm 16 0 1.017527\n",
" 8 0 1.012077\n",
"HCP spm 16 0 1.007503\n",
" fsl 1 6 1.005294\n",
"ds001545 afni 16 0 1.002235\n",
"HCP spm 8 0 0.995598\n",
"preventAD spm 16 0 0.993274\n",
"HCP afni 16 0 0.992692\n",
"ds001545 fsl 16 6 0.986928\n",
"preventAD fsl 1 0 0.986418\n",
"HCP fsl 8 6 0.983173\n",
"ds001545 fsl 1 0 0.981228\n",
" afni 8 0 0.976202\n",
"HCP fsl 16 6 0.971510\n",
"preventAD afni 8 0 0.956270\n",
"HCP fsl 1 0 0.945483\n",
"preventAD fsl 8 0 0.934627\n",
"HCP fsl 8 0 0.934242\n",
"ds001545 fsl 8 0 0.919897\n",
"preventAD afni 16 6 0.913218\n",
"HCP fsl 16 0 0.896197\n",
"preventAD fsl 16 0 0.891131\n",
"ds001545 fsl 16 0 0.866530\n",
"preventAD spm 8 6 0.712436\n",
"ds001545 afni 8 6 0.638075\n",
"Name: speed_up, dtype: float64"
]
},
"execution_count": 47,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"df_merged.groupby(by=['dataset', 'pipeline', 'n_images', 'busy_writers'])['speed_up'].mean().sort_values(ascending=False)"
]
},
{
"cell_type": "markdown",
"id": "f6568ab1",
Expand Down

0 comments on commit 552b63b

Please sign in to comment.