Recent advances in Multimodal Large Language Models (MLLMs) have spurred significant progress in Chain-of-Thought (CoT) reasoning. Building on the success of Deepseek-R1, researchers extended multimodal reasoning to post-training paradigms based on reinforcement learning (RL), focusing predominantly on mathematical datasets. However, existing post-training paradigms tend to neglect two critical aspects: (1) The lack of quantifiable difficulty metrics capable of strategically screening samples for post-training optimization. (2) Suboptimal post-training paradigms that fail to jointly optimize perception and reasoning capabilities. To address this gap, we propose two novel difficulty-aware sampling strategies: Progressive Image Semantic Masking (PISM) quantifies sample hardness through systematic image degradation, while Cross-Modality Attention Balance (CMAB) assesses cross-modal interaction complexity via attention distribution analysis. Leveraging these metrics, we design a hierarchical training framework that incorporates both GRPO-only and SFT+GRPO hybrid training paradigms, and evaluate them across six benchmark datasets. Experiments demonstrate consistent superiority of GRPO applied to difficulty-stratified samples compared to conventional SFT+GRPO pipelines, indicating that strategic data sampling can obviate the need for supervised fine-tuning while improving model accuracy.

The parameters of deep learning (DL) have ballooned from millions to trillions over the past decade, thus cannot be fully placed in a limited GPU memory. Existing works offload the parameter-update stage of training DL model to CPU, thus leveraging CPU capability and memory to support training large-scale model. However, the stage running upon CPU could block its following stage running upon GPU, thus wasting expensive GPU-cycles. We first analyze the dataflow and workflow of DL training, and find that the backward stage and the parameter-update stage can be parallelized upon GPU and CPU respectively. To this end, we present a DL-training scheduling framework, CoTrain, to allocate a compute task and its corresponding data into GPU and CPU and to parallelize them effectively at coarse-grain and fine-grain way. Particularly, the fine-grained task-partition scheme allocates a portion of parameter-update stage to GPU according to data-reuse-distance, thus largely avoiding idleness of both GPU and CPU while reducing data movement between GPU and CPU. We build and evaluate CoTrain atop PyTorch under representative models. The results show that compared to the state-of-the-art ZeRO-Offload, CoTrain achieves 30.4% improvement in the training throughput while increasing model-size by up to 7% without changing the training semantics.

Data centers have been becoming public utilities to provide large-scale computing and storage services. The Total Cost of Ownership (TCO) models for such data centers are paramount to deeply understand their cost of investment and maintenance, the cost composition of internal components, and further cost optimization directions. Existing data center TCO models focus on either high-performance data centers or key subsystems such as IT facility, lacking of holistic analysis of the data centers designed for long-term data preservation. The long-term data centers can be built with different combinations of storage media such as HDDs, tapes, and optical discs. Meanwhile, during the long operation period, devices replacement and data migration are necessary and are not negligible in cost. In order to comprehensively and quantitatively understand the cost of long-term data preservations, we proposed LT-TCO, a TCO calculation model for data centers over time. LT-TCO simulates the construction and operation of a data center to calculate the expenditure of each year. It also introduces the cost of devices replacement and data migration during the long running period. Based on the storage media as optical discs, tapes, HDDs, and SSDs, LT-TCO evaluates the corresponding capital and operational expenditure under different developing rates. The simulation result shows that in long-term preservation, data migration cost takes more than 96% of the operational expenditure. And the TCO of optical disc data centers could be the least among four storage media.

The combination of the explosive growth in digital data and the demand to preserve much of these data in the long term has made it imperative to find a more cost-effective way than HDD arrays and a more easily accessible way than tape libraries to store massive amounts of data. While modern optical discs are capable of guaranteeing more than 50-year data preservation without media replacement, individual optical discs’ lack of the performance and capacity relative to HDDs or tapes has significantly limited their use in datacenters. This article presents a Rack-scale Optical disc library System, or ROS in short, which provides a PB-level total capacity and inline accessibility on thousands of optical discs built within a 42U Rack. A rotatable roller and robotic arm separating and fetching discs are designed to improve disc placement density and simplify the mechanical structure. A hierarchical storage system based on SSDs, hard disks, and optical discs is proposed to effectively hide the delay of mechanical operation. However, an optical library file system (OLFS) based on FUSE is proposed to schedule mechanical operation and organize data on the tiered storage with a POSIX user interface to provide an illusion of inline data accessibility. We further optimize OLFS by reducing unnecessary user/kernel context switches inheriting from legacy FUSE framework. We evaluate ROS on a few key performance metrics, including operation delays of the mechanical structure and software overhead in a prototype PB-level ROS system. The results show that ROS stacked on Samba and FUSE as network-attached storage (NAS) mode almost saturates the throughput provided by underlying samba via 10GbE network for external users, as well as in this scenario provides about 53ms file write and 15ms read latency, exhibiting its inline accessibility. Besides, ROS is able to effectively hide and virtualize internal complex operational behaviors and be easily deployable in datacenters.

The combination of the explosive growth in digital data and the need to preserve much of this data in the long term has made it an imperative to find a more cost-effective way than HDD arrays and more easily accessible way than tape libraries to store massive amounts of data. While modern optical discs are capable of guaranteeing more than 50-year data preservation without migration, individual optical disks' lack of the performance and capacity relative to HDDs or tapes has significantly limited their use in datacenters. This paper presents a Rack-scale Optical disc library System, or ROS in short, that provides a PB-level total capacity and inline accessibility on thousands of optical discs built within a 42U Rack. A rotatable roller and robotic arm separating and fetching the discs are designed to improve disc placement density and simplify the mechanical structure. A hierarchical storage system based on SSD, hard disks and optical discs are presented to hide the delay of mechanical operation. On the other hand, an optical library file system is proposed to schedule mechanical operation and organize data on the tiered storage with a POSIX user interface to provide an illusion of inline data accessibility. We evaluate ROS on a few key performance metrics including operation delays of the mechanical structure and software overhead in a prototype PB-level ROS system. The results show that ROS stacked on Samba and FUSE can provide almost 323MB/s read and 236MB/s write throughput, about 53ms file write and 15ms read latency via 10GbE network for external users, exhibiting its inline accessibility. Besides, ROS is able to effectively hide and virtualize internal complex operational behaviors and be easily deployable in datacenters.