[Computer Architecture]

In this work, I implemented support for two page sizes (4KB and 64KB) in the GPGPU-simulator to analyze the impact of large pages on application performance in discrete GPUs. I evaluated the performance of this support using microbenchmarks and several standard benchmarks, comparing it against the performance achieved with only 4KB pages. The results showed that using large pages improved Instructions Per Cycle (IPC) for most applications while reducing the number of page walks and the L2 TLB miss rate. Additionally, for multi-tenancy workloads, IPC improved significantly.

[Computer Graphics and Visualization]

In this work, I visualized high-dimensional scientific data using a low-dimensional topological landscape profile. The input model was represented as a join tree, which encodes geometric information about points along its arcs. I leveraged this information to construct both 1D and 2D topological landscape profiles, providing insights into the structure of the data.

The figure on the left illustrates the cat model with isovalues mapped onto its surface, where red represents low isovalues and blue represents high isovalues. The accompanying images depict the 1D and 2D landscape profiles derived from the model. In the 1D landscape profile, six distinct hills correspond to six high-isovalue regions. Notably, the leftmost hill, which exhibits the highest persistence, corresponds to the tail of the cat model. The four narrow hills represent the cat’s legs, while the broad hill corresponds to the cat’s head. Some hills are not entirely blue, which may indicate that the edges they represent have further children, potentially due to noise or smaller structural components within the model.

[Computer Graphics and Visualization]

Analyzing dynamic networks is inherently complex, as it requires incorporating time-varying dissimilarities into the analysis. While Topological Data Analysis (TDA) has been extensively studied for static networks and point clouds, I explored its application to time-varying networks to extract meaningful insights, such as the formation of cycles, one-time events, and other temporal patterns.

In this work, I analyzed the high school dynamic contact dataset, which captures the temporal network of contacts between 180 students across five classes over seven school days. The graphs on the left illustrate the analysis of a single day. Initially, there is no recorded activity at the start of the day. Activity begins around 4:54 AM and steadily increases until 7:00 AM. After 7:06 AM, a significant rise in interactions leads to the formation of larger clusters. This heightened activity continues until 8:00 AM, after which there is a noticeable drop in the number of interactions, as reflected in the timeline.