Reorganize image files

Author: Edoardo Holzl

_drafts/Benchmark1-scaling.md

@@ -101,12 +101,12 @@ The definitions helm chart values can be found [here](https://mlbench.readthedoc
 
 ## Results
 * Epochs to Top-1 Validation Accuracy
-<a href="{{ site.baseurl }}public/images/scaling-epoch-prec1.png" data-lightbox="Run" data-title="Validation Accuracy @ 1">
-  <img src="{{ site.baseurl }}public/images/scaling-epoch-prec1.png" alt="Validation Accuracy @ 1" style="max-width:80%;"/>
+<a href="{{ site.baseurl }}public/images/blog/drafts/scaling-epoch-prec1.png" data-lightbox="Run" data-title="Validation Accuracy @ 1">
+  <img src="{{ site.baseurl }}public/images/blog/drafts/scaling-epoch-prec1.png" alt="Validation Accuracy @ 1" style="max-width:80%;"/>
 </a>
 * Time to Top-1 Validation Accuracy
-<a href="{{ site.baseurl }}public/images/scaling-time-prec1.png" data-lightbox="Run" data-title="Validation Accuracy @ 1">
-  <img src="{{ site.baseurl }}public/images/scaling-time-prec1.png" alt="Validation Accuracy @ 1" style="max-width:80%;"/>
+<a href="{{ site.baseurl }}public/images/blog/drafts/scaling-time-prec1.png" data-lightbox="Run" data-title="Validation Accuracy @ 1">
+  <img src="{{ site.baseurl }}public/images/blog/drafts/scaling-time-prec1.png" alt="Validation Accuracy @ 1" style="max-width:80%;"/>
 </a>
 
 

_drafts/MLBench-benchmarking-mpi-speed.md

@@ -17,8 +17,8 @@ In this experiment, we compare MPI P2P communication for
 CUDA Inter-Process Communication (IPC) improves communication between GPUs on the same node. In openmpi, one can use `--mca btl_smcuda_use_cuda_ipc` to turn on/off this functionality. We demostrate the influence of CUDA-IPC by sending/receiving a vector on a node with two GPUs. 
 
 ### Results
-<a href="{{ site.baseurl }}public/images/mpi-speed-p2p.png" data-lightbox="Run" data-title="MPI Speed P2P">
-  <img src="{{ site.baseurl }}public/images/mpi-speed-p2p.png" alt="MPI Speed P2P" style="max-width:80%;"/>
+<a href="{{ site.baseurl }}public/images/blog/drafts/mpi-speed-p2p.png" data-lightbox="Run" data-title="MPI Speed P2P">
+  <img src="{{ site.baseurl }}public/images/blog/drafts/mpi-speed-p2p.png" alt="MPI Speed P2P" style="max-width:80%;"/>
 </a>
 
 - The P2P communication between two nodes is bounded by network bandwidth (`7.5 Gbit/s` measured by `iperf`). Communicating large vectors on CPU/GPU have similar throughput.
@@ -41,8 +41,8 @@ The connection between GPUs are PHB which traverss PCIe as well as a PCIe Host B
 ### Results
 The results of experiments are shown below. Note that the bandwidth here is calculated by dividing the vector size by the time it spent. The actual bandwidth depends on the implementation of all reduce.
 
-<a href="{{ site.baseurl }}public/images/mpi-speed-collective.png" data-lightbox="Run" data-title="MPI Speed Collective">
-  <img src="{{ site.baseurl }}public/images/mpi-speed-collective.png" alt="MPI Speed Collective" style="max-width:80%;"/>
+<a href="{{ site.baseurl }}public/images/blog/drafts/mpi-speed-collective.png" data-lightbox="Run" data-title="MPI Speed Collective">
+  <img src="{{ site.baseurl }}public/images/blog/drafts/mpi-speed-collective.png" alt="MPI Speed Collective" style="max-width:80%;"/>
 </a>
 
 - The NCCL all reduce does not give better performance when the GPU per machine is 1 or 2.

_drafts/MLBench-limits-lie-00.md

@@ -15,22 +15,22 @@ First create a cluster of two `n1-standard-4` instances with `limits.cpu=1000m`
 
 ### master/worker-0 node
 
-<a href="{{ site.baseurl }}public/images/lie-00-resource-node1.png" data-lightbox="Dashboard_Index" data-title="lie-00-resource-node1">
-  <img src="{{ site.baseurl }}public/images/lie-00-resource-node1.png" alt="lie-00-resource-node1" style="max-width:100%;"/>
+<a href="{{ site.baseurl }}public/images/blog/drafts/lie-00-resource-node1.png" data-lightbox="Dashboard_Index" data-title="lie-00-resource-node1">
+  <img src="{{ site.baseurl }}public/images/blog/drafts/lie-00-resource-node1.png" alt="lie-00-resource-node1" style="max-width:100%;"/>
 </a>
 
 Only 2 pods are for mlbench: `release1-mlbench-master-6448bfb454-sxm2l` (`100m` CPU) and `release1-mlbench-worker-0` (`1000m` CPU). The rest of pods request `1161m` of CPU and `750MB` memory.
 
 The summary of resources on this node is (requests `2261m` CPU in total )
 
-<a href="{{ site.baseurl }}public/images/lie-00-resource-node1-summary.png" data-lightbox="Dashboard_Index" data-title="The MLBench Dashboard">
-  <img src="{{ site.baseurl }}public/images/lie-00-resource-node1-summary.png" alt="The MLBench Dashboard" style="max-width:100%;"/>
+<a href="{{ site.baseurl }}public/images/blog/drafts/lie-00-resource-node1-summary.png" data-lightbox="Dashboard_Index" data-title="The MLBench Dashboard">
+  <img src="{{ site.baseurl }}public/images/blog/drafts/lie-00-resource-node1-summary.png" alt="The MLBench Dashboard" style="max-width:100%;"/>
 </a>
 
 ### worker-1 node
 On worker-1 node, there are much less pods.
-<a href="{{ site.baseurl }}public/images/lie-00-resource-node2.png" data-lightbox="Dashboard_Index" data-title="lie-00-resource-node2">
-  <img src="{{ site.baseurl }}public/images/lie-00-resource-node2.png" alt="lie-00-resource-node2" style="max-width:100%;"/>
+<a href="{{ site.baseurl }}public/images/blog/drafts/lie-00-resource-node2.png" data-lightbox="Dashboard_Index" data-title="lie-00-resource-node2">
+  <img src="{{ site.baseurl }}public/images/blog/drafts/lie-00-resource-node2.png" alt="lie-00-resource-node2" style="max-width:100%;"/>
 </a>
 
 So the amount of resources available is limited to the master node. In the previous setting we can allocate at most `3920-1161-100=2659m` for each worker.

_posts/2018-09-07-introducing-mlbench.md

@@ -8,8 +8,8 @@ excerpt_separator: <!--more-->
 ---
 MLBench is a framework for distributed machine learning. Its purpose is to improve transparency, reproducibility, robustness, and to provide fair performance measures as well as reference implementations, helping adoption of distributed machine learning methods both in industry and in the academic community.
 
-<a href="{{ site.baseurl }}public/images/Dashboard_Index.png" data-lightbox="Dashboard_Index" data-title="The MLBench Dashboard">
-  <img src="{{ site.baseurl }}public/images/Dashboard_Index.png" alt="The MLBench Dashboard" style="max-width:80%;"/>
+<a href="{{ site.baseurl }}public/images/blog/2018-09-07-introducing-mlbench/Dashboard_Index.png" data-lightbox="Dashboard_Index" data-title="The MLBench Dashboard">
+  <img src="{{ site.baseurl }}public/images/blog/2018-09-07-introducing-mlbench/Dashboard_Index.png" alt="The MLBench Dashboard" style="max-width:80%;"/>
 </a>
 
 <!--more-->

_posts/2020-09-08-communication-backend-comparison.md

@@ -73,8 +73,8 @@ and can be sped up using distributed training.
 #### CPU
 In the graph below, we compare the speeds taken to perform an `all reduce` operation between 2, 4 and 8 workers, of `Float16` and `Float32` CPU tensors.
 
-<a href="{{ site.baseurl }}public/images/backends_comparison_by_workers.png" data-lightbox="backends_comparison_by_workers" data-title="Backend performance comparison (CPU tensors)">
-  <img src="{{ site.baseurl }}public/images/backends_comparison_by_workers.png" alt="Backend performance comparison (CPU tensors)" style="max-width:100%;"/>
+<a href="{{ site.baseurl }}public/images/blog/2020-09-08-communication-backend-comparison/backends_comparison_by_workers.png" data-lightbox="backends_comparison_by_workers" data-title="Backend performance comparison (CPU tensors)">
+  <img src="{{ site.baseurl }}public/images/blog/2020-09-08-communication-backend-comparison/backends_comparison_by_workers.png" alt="Backend performance comparison (CPU tensors)" style="max-width:100%;"/>
 </a>
 
 ##### Key differences
@@ -88,8 +88,8 @@ In the graph below, we compare the speeds taken to perform an `all reduce` opera
 
 We now compare the speeds for GPU tensors. Here, we have the addition of NCCL in the comparison.
 
-<a href="{{ site.baseurl }}public/images/backends_comparison_by_workers_CUDA.png" data-lightbox="backends_comparison_by_workers" data-title="Backend performance comparison (GPU tensors)">
-  <img src="{{ site.baseurl }}public/images/backends_comparison_by_workers_CUDA.png" alt="Backend performance comparison (GPU tensors)" style="max-width:100%;"/>
+<a href="{{ site.baseurl }}public/images/blog/2020-09-08-communication-backend-comparison/backends_comparison_by_workers_CUDA.png" data-lightbox="backends_comparison_by_workers" data-title="Backend performance comparison (GPU tensors)">
+  <img src="{{ site.baseurl }}public/images/blog/2020-09-08-communication-backend-comparison/backends_comparison_by_workers_CUDA.png" alt="Backend performance comparison (GPU tensors)" style="max-width:100%;"/>
 </a>
 
 ##### Key differences

index.md

@@ -66,10 +66,10 @@ Check out our <a href="https://mlbench.github.io/blog/">blog</a>!
 <h2>Sponsors</h2>
 
 <ul style="list-style-type:none;">
-    <li><img src="{{ site.baseurl }}public/images/Logo_EPFL.png" alt="EPFL" style="max-width:200px;border-radius:0px;"/></li>
-    <li><img src="{{ site.baseurl }}public/images/pwc_logo.png" alt="PwC" style="max-width:200px;border-radius:0px;"/></li>
-    <li><img src="{{ site.baseurl }}public/images/google.png" alt="Google" style="max-width:200px;border-radius:0px;"/></li>
-    <li><img src="{{ site.baseurl }}public/images/Facebook-Wordmark-Gray.png" alt="Facebook" style="max-width:200px;border-radius:0px;"/></li>
+    <li><img src="{{ site.baseurl }}public/images/assets/Logo_EPFL.png" alt="EPFL" style="max-width:200px;border-radius:0px;"/></li>
+    <li><img src="{{ site.baseurl }}public/images/assets/pwc_logo.png" alt="PwC" style="max-width:200px;border-radius:0px;"/></li>
+    <li><img src="{{ site.baseurl }}public/images/assets/google.png" alt="Google" style="max-width:200px;border-radius:0px;"/></li>
+    <li><img src="{{ site.baseurl }}public/images/assets/Facebook-Wordmark-Gray.png" alt="Facebook" style="max-width:200px;border-radius:0px;"/></li>
 </ul>
 
 

public/css/lightbox.css

@@ -76,7 +76,7 @@ html.lb-disable-scrolling {
   width: 32px;
   height: 32px;
   margin: 0 auto;
-  background: url(../images/loading.gif) no-repeat;
+  background: url(../images/assets/loading.gif) no-repeat;
 }
 
 .lb-nav {
@@ -107,7 +107,7 @@ html.lb-disable-scrolling {
   width: 34%;
   left: 0;
   float: left;
-  background: url(../images/prev.png) left 48% no-repeat;
+  background: url(../images/assets/prev.png) left 48% no-repeat;
   filter: progid:DXImageTransform.Microsoft.Alpha(Opacity=0);
   opacity: 0;
   -webkit-transition: opacity 0.6s;
@@ -125,7 +125,7 @@ html.lb-disable-scrolling {
   width: 64%;
   right: 0;
   float: right;
-  background: url(../images/next.png) right 48% no-repeat;
+  background: url(../images/assets/next.png) right 48% no-repeat;
   filter: progid:DXImageTransform.Microsoft.Alpha(Opacity=0);
   opacity: 0;
   -webkit-transition: opacity 0.6s;
@@ -189,7 +189,7 @@ html.lb-disable-scrolling {
   float: right;
   width: 30px;
   height: 30px;
-  background: url(../images/close.png) top right no-repeat;
+  background: url(../images/assets/close.png) top right no-repeat;
   text-align: right;
   outline: none;
   filter: progid:DXImageTransform.Microsoft.Alpha(Opacity=70);