Warehouse-scale video acceleration: co-design and deployment in the wild

Ranganathan, Parthasarathy; Stodolsky, Daniel; Calow, Jeff; Dorfman, Jeremy; Guevara, Marisabel; Smullen, Clinton W.; Kuusela, Aki; Balasubramanian, Raghu; Bhatia, Sandeep; Chauhan, Prakash; Cheung, Anna Wing Han; Chong, In Suk; Dasharathi, Niranjani; Feng, Jingyu; Fosco, Brian; Foss, Samuel; Gelb, Ben; Gwin, Sara J.; Hase, Yoshiaki; He, Dongxiao; Ho, Chien-Te; Huffman, Roy W.; Indupalli, Elisha; Jayaram, Indira; Kongetira, Poonacha; Kyaw, Cho Mon; Laursen, Aaron; Li, Yuan; Lou, Fong; Lucke, Kyle; Maaninen, JP; Macías, R Medina; Mahony, Maire; Munday, David Alexander; Muroor, Srikanth; Penukonda, Narayana; Perkins-Argueta, Eric; Persaud, Devin; Ramírez, Alex; Rautio, Ville-Mikko; Ripley, Yolanda; Salek, Amir; Sekar, Sathish; Sokolov, S. N.; Springer, Rob; Stark, D.; Tan, Mercedes; Wachsler, Mark; Walton, Amie; Wickeraad, David A.; Wijaya, Alvin; Wu, Hon Kwan

doi:10.1145/3445814.3446723

Cited by 34 publications

(7 citation statements)

References 46 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…A second example further illustrates “hardware-software co-design” (Ranganathan et al, 2021), that is, the close coordination between hardware and software components around specific application requirements. To deal with the vast number of uploaded videos and the increasing diversity of viewing devices, YouTube has designed Video Coding Units (VCUs) that can transcode files to different formats in parallel, using a “multiple-output transcoding” approach where certain processing steps are shared to increase throughput and efficiency (Ranganathan et al, 2021). Google also uses these VCUs for their cloud gaming service Stadia, again showing how product variety enables synergies that make costly investments in cutting edge technologies viable.…”

Section: Google As Technical Systemmentioning

confidence: 99%

Towards a political economy of technical systems: The case of Google

Rieder

2022

Big Data & Society

View full text Add to dashboard Cite

This research commentary proposes a conceptual framework for studying big tech companies as “technical systems” that organize much of their operation around the mastery and operationalization of key technologies that facilitate and drive their continuous expansion. Drawing on the study of Large Technical Systems (LTS), on the work of historian Bertrand Gille, and on the economics of General Purpose Technologies (GPTs), it outlines a way to study the “tech” in “big tech” more attentively, looking for compatibilities, synergies, and dependencies between the technologies created and deployed by these companies. Using Google as example, the paper shows how to interrogate software and hardware through the lens of transversal applicability, discusses software and hardware integration, and proposes the notion of “data amalgams” to contextualize and complicate the notion of data. The goal is to complement existing vectors of “big tech” critique with a perspective sensitive to the specific materialities of specific technologies and their possible consequences.

show abstract

Section: Google As Technical Systemmentioning

confidence: 99%

Towards a political economy of technical systems: The case of Google

Rieder

2022

Big Data & Society

View full text Add to dashboard Cite

show abstract

“…Several cloud providers already integrate ASIC accelerators in their datacenters. For example, Google Cloud, Microsoft Azure and Amazon AWS use dedicated hardware for neural network inference (Google Cloud TPU [184,185], Habana Goya [186]), neural network training (AWS Trainium [187], Habana Gaudi [186]), video transcoding (Google VCU [188]), and compression/encryption/data authentication (Microsoft's Project Corsica [189]). Recently, different commercial PIM designs [95-98, 108-111, 190], which target large cloud systems, have been proposed.…”

Section: Integrating Polynesia In the Cloudmentioning

confidence: 99%

“…Polynesia's energy savings ( §10.6) are also attractive to a cloud environment, as it can significantly lower operating costs. As prior works show [188,191], ASIC accelerators are a viable solution for cloud environments, depending on the scale of the computation. Even though we cannot accurately predict the price of integrating Polynesia into a cloud system due to unknown parameters (e.g., non recurring engineering expenses), we believe it would be a beneficial solution for cloud systems due to the widespread use of the database applications it targets.…”

Section: Integrating Polynesia In the Cloudmentioning

confidence: 99%

Enabling High-Performance and Energy-Efficient Hybrid Transactional/Analytical Databases with Hardware/Software Cooperation

Boroumand¹,

Ghose²,

F.³

et al. 2022

Preprint

View full text Add to dashboard Cite

A growth in data volume, combined with increasing demand for real-time analysis (using the most recent data), has resulted in the emergence of database systems that concurrently support transactions and data analytics. These hybrid transactional and analytical processing (HTAP) database systems can support real-time data analysis without the high costs of synchronizing across separate single-purpose databases. Unfortunately, for many applications that perform a high rate of data updates, state-of-the-art HTAP systems incur significant losses in transactional (up to 74.6%) and/or analytical (up to 49.8%) throughput compared to performing only transactional or only analytical queries in isolation, due to (1) data movement between the CPU and memory, (2) data update propagation from transactional to analytical workloads, and (3) the cost to maintain a consistent view of data across the system.We propose Polynesia, a hardware-software co-designed system for in-memory HTAP databases that avoids the large throughput losses of traditional HTAP systems. Polynesia (1) divides the HTAP system into transactional and analytical processing islands, (2) implements new custom hardware that unlocks software optimizations to reduce the costs of update propagation and consistency, and (3) exploits processing-in-memory for the analytical islands to alleviate data movement overheads. Our evaluation shows that Polynesia outperforms three stateof-the-art HTAP systems, with average transactional/analytical throughput improvements of 1.7×/3.7×, and reduces energy consumption by 48% over the prior lowest-energy HTAP system.

show abstract

“…Because the fixed function hardware cores are nonfungible, we moved from a uniform CPU cost scheduler to an online multidimensional bin packing scheduler [shown in Sec. 3.3.3 of the work by Ranganathan et al 3 ] to ensure no single VCU becomes completely saturated and no encoder cores become starved.…”

Section: Cluster and Warehouse-scale Schedulingmentioning

confidence: 99%