4ms-400
Wtf is 400ms? We need more insights about the dataset
7 Matching Annotations
- Last 7 days
-
Local file Local file
-
-
The paper includes a real Washington–Utah deployment only for the decision-delay sensitivity study, while the main Orca and Pensieve performance claims rely on trace-based emulation. It is therefore unclear whether the reported end-to-end gains carry over to a live Internet deployment.
-
The paper motivates UNUM as capturing latent factors such as workload, cross traffic, and controller interactions, but most embedding analysis demonstrates separation by explicit network parameters such as bandwidth and RTT rather than directly validating that these harder latent factors are encoded.
+
they talk about shared bottleneck etc but they dont address it?
-
UNUM's main motivation is generalization, yet the training corpus is described primarily by parameter ranges and categories rather than by the empirical distribution of network conditions. It is therefore difficult to tell what kinds of environments dominate the learned representation or how far the evaluation actually shifts from the training distribution.
UNUM emphasizes cross-environment generalization, but the paper reports only parameter ranges for its training corpus rather than the distribution of network conditions. The released CC code appears to systematically sweep a bandwidth–RTT grid for synthetic training while evaluating selected Mahimahi/Pantheon-derived trace sets. How sensitive is UNUM to the training distribution, and how does its performance change when the test distribution over RTT, capacity dynamics, and queueing behavior differs substantially from the synthetic training distribution?
-
The paper's headline gains rely heavily on controller-specific reward functions, which hide the tradeoffs between the actual end-to-end metrics being optimized..
The paper relies too much on controller-specific test reward to summarize gains. Since reward functions already encode tradeoffs between metrics such as throughput, delay, rebuffering, and quality variation, a large reward improvement does not necessarily mean an equally large real QoE improvement. For example, Orca-UNUM improves utilization but slightly increases queueing delay, while Pensieve’s reward includes quality variation that is not separately shown. The BBR results are more convincing because they report direct changes in utilization and delay.
-
Decision delay breakdown. We evaluate the impact of keyembedder model hyperparameters on decision delay. Amongthem, only the number of encoder layers and embedding sizehave an observable influence. Figure 16a and Figure 16bshow that the dominant contributor to decision delay is theinference time. Inference delay ranges from 2.2 ms to 23ms as the number of encoder layers increases from 2 to 32,and from 2.2 ms to 3.2 ms as the embedding size increasesfrom 8 to 256. For the default UNUM embedder configuration(4 encoder layers, 16 embedding size), the average decisiondelay is 2.68 ms (data collection: 38.95 ns, data movement:0.12 ms, tokenization: 0.01 ms, inference: 2.44 ms). Sinceboth Orca and Pensieve operate at coarse-grained timescalesfor control decisions, such a small added delay is acceptable
I would like to see inference latency vs. improvement graph
-
3kB - 96MB
Again the distribution not clear.
-