Skip to content

Commit d13ac28

Browse files
authored
Plot both PDF and PNG figures (#2314)
1 parent 19ebd7c commit d13ac28

10 files changed

+6
-6
lines changed
-8 Bytes
Binary file not shown.
36.3 KB
Loading
-14 Bytes
Binary file not shown.
31.2 KB
Loading
1 Byte
Binary file not shown.
114 KB
Loading

docs/benchmark/allreduce/report.md

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -58,7 +58,7 @@ The following figure compares the gang (non-elastic) and elastic scheduling
5858
using ElasticDL. The Kubernetes cluster has 6 GPUs, and each of the two jobs
5959
requests 4 GPUs.
6060

61-
![overlap jobs](./data/experiment_1.pdf)
61+
![overlap jobs](./data/experiment_1.png)
6262

6363
The upper part of the figure presents the GPU utilization over time without
6464
elastic scheduling. We see that the two jobs run one after another. Between
@@ -85,7 +85,7 @@ To use the idle resource, we can run ElasticDL training jobs with lower
8585
priorities than the inference service. To verify the idea, we run the
8686
following experiment.
8787

88-
![preemption](./data/experiment_2.pdf)
88+
![preemption](./data/experiment_2.png)
8989

9090
This experiment uses a Kubernetes cluster with 8 GPUs. We start a TensorFlow
9191
Serving service and an ElasticDL training job — a script program mimics user
@@ -121,7 +121,7 @@ the CIFAR10 dataset. Each worker uses an NVIDIA Tesla P100 GPU. Curves marked
121121
"baseline" corresponds to experiments using gang scheduling, and the mark
122122
"elastic" denotes ElasticDL jobs.
123123

124-
![accuracy](./data/experiment_3.pdf)
124+
![accuracy](./data/experiment_3.png)
125125

126126
ElasticDL allows users to define a function of learning rate decay. This
127127
experiment uses the following decay function.

docs/benchmark/allreduce/scripts/acc_plot.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -45,5 +45,5 @@
4545
plt.ylabel(ylabel="Accuracy", fontsize=18, fontfamily="Times New Roman")
4646
plt.legend(loc="upper left", bbox_to_anchor=(0.02, 0.98))
4747
plt.ylim((0.3, 1))
48-
plt.show()
4948
f.savefig("../data/experiment_3.pdf", bbox_inches="tight")
49+
f.savefig("../data/experiment_3.png", bbox_inches="tight")

docs/benchmark/allreduce/scripts/gpu_uilization_plot.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -43,5 +43,5 @@
4343
plt.xlim(0, 1000)
4444
plt.ylim(0, 7)
4545
plt.subplots_adjust(hspace=0.5)
46-
plt.show()
4746
fig.savefig("../data/experiment_1.pdf", bbox_inches="tight")
47+
fig.savefig("../data/experiment_1.png", bbox_inches="tight")

docs/benchmark/allreduce/scripts/preempt_gpu_plot.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -38,5 +38,5 @@
3838
plt.xlim(0, 500)
3939
plt.ylim(0, 11)
4040
plt.subplots_adjust(hspace=0.5)
41-
plt.show()
4241
f.savefig("../data/experiment_2.pdf", bbox_inches="tight")
42+
f.savefig("../data/experiment_2.png", bbox_inches="tight")

0 commit comments

Comments
 (0)