YOSHIDA Hideki
LightNews LightNews
teikan.bsky.social
YOSHIDA Hideki
@teikan.bsky.social
21 followers 52 following 92 posts
sync; sync; sync; umount
@teikan
@[email protected]
Posts Replies Media Videos
teikan.bsky.social
普通のサーバの償却期間は前から伸びてますが、高性能GPUで5年以上はさすがに無理がある
xenospectrum.com/michael-burr...
「世紀の空売り」男マイケル・バーリがファンドを閉鎖：AIへの反逆は失敗か？ | XenoSpectrum
2008年の世界金融危機を正確に予測し、映画『マネー・ショート 華麗なる大逆転』でその名を世界に轟かせた伝説の投資家、Michael
xenospectrum.com
November 15, 2025 at 9:50 PM
teikan.bsky.social
来週PDSW '25
www.pdsw.org

これが気になる
SlimIO: Lightweight I/O Path Design for Write Isolation in FDP-backed In-Memory Databases
dl.acm.org/doi/10.1145/...
Complimentary by FCT
www.pdsw.org
November 11, 2025 at 7:32 AM
teikan.bsky.social
ARM移行じゃなくてマルチISA対応なところがすごい
October 22, 2025 at 9:26 AM
teikan.bsky.social
物理破壊廃止
cloud.google.com/blog/product...
The future of media sanitization at Google | Google Cloud Blog
To help you get started in pursuing malware analysis as a primary specialty, we’re introducing Mandiant Academy’s new "Basic Static and Dynamic Analysis" course.
cloud.google.com
October 13, 2025 at 9:56 PM
teikan.bsky.social
APSys '25論文。Linuxカーネルにもう入ったXFSのZNS対応。lifetime hintが効いているらしい
dl.acm.org/doi/abs/10.1...
Staying in the Zone - Retrofitting Zoned Storage into a Scalable Enterprise File System | Proceedings of the 16th ACM SIGOPS Asia-Pacific Workshop on Systems
dl.acm.org
October 13, 2025 at 8:33 AM
teikan.bsky.social
大型投資・買収のニュースにまぎれてますが、Nitro/Trainium/InferentiaのAnnapurnaの人。以前もArmにいたらしい
August 19, 2025 at 7:54 AM
teikan.bsky.social
再来週のHotStorageのAccepted Papers
www.hotstorage.org/2025/accepte...
HOTSTORAGE 2025
HOTSTORAGE'25: 17th ACM Workshop on Hot Topics in Storage and File Systems
www.hotstorage.org
June 26, 2025 at 9:16 PM
teikan.bsky.social
ハイパースケーラーはまだドライブ物理破壊してるらしい。
従業員が勝手に増設したドライブが見つかるというのも面白い
May 26, 2025 at 11:19 PM
teikan.bsky.social
処理能力が高く省電力な次世代メモリとは
k-tai.watch.impress.co.jp/docs/news/20...
ソフトバンクがAIデータセンター向けの“次世代メモリー”開発に着手、2年間で約30億円を投資
ソフトバンクは8日、同社の2025年3月期 決算説明会で、AIデータセンター向けの“次世代メモリー”の開発に着手すると発表した。製造ではなく、IP（知的財産、ライセンス）事業を想定している。
k-tai.watch.impress.co.jp
May 8, 2025 at 8:22 AM
teikan.bsky.social
SMRの管理方式の話いろいろ。非SMR領域を真ん中にするとシークが減るとか
dl.acm.org/doi/10.1145/...
Reflecting on the Past 17 Years of Shingled Magnetic Recording for Insights into Future Disk Transitions: A Survey | ACM Transactions on Storage
Shingled magnetic recording (SMR) is a data storage recording technology used in modern hard disk drives (HDDs) to increase the areal density capacity (ADC) of underlying media. The research on SMR dr...
dl.acm.org
April 26, 2025 at 9:05 PM
teikan.bsky.social
handleやdurable appendがCloud Storage FUSEで利用可能に
April 10, 2025 at 10:06 PM
teikan.bsky.social
回転する前提でシステム設計
thessdguy.com/why-ssds-are...
Why SSDs are Problematic in Space – The SSD Guy Blog
thessdguy.com
April 1, 2025 at 9:30 PM
teikan.bsky.social
OSDI 2025採択
www.csail.mit.edu/event/tigon-...
Tigon: A Distributed Transactional Database for a CXL Pod | MIT CSAIL
www.csail.mit.edu
March 30, 2025 at 9:02 PM
teikan.bsky.social
CacheSack方式をwrite backキャッシュにも応用
March 26, 2025 at 8:54 PM
teikan.bsky.social
RAGとは違って外部データから直接アテンション作ってLLMに入れる。省資源でハルシネーション少ないそう
msftresearch.bsky.social Microsoft Research @msftresearch.bsky.social · Mar 18
Introducing KBLaM, an approach that encodes and stores structured knowledge within an LLM itself. By integrating knowledge without retraining, it offers a scalable alternative to traditional methods. www.microsoft.com/en-us/resear...
A flowchart illustrating the process of handling a prompt using a language model. The process begins with documents being used to construct and summarize a knowledge base (KB) offline. The summarized KB is then encoded and fed into the main process. A prompt goes through a tokenizer, followed by rectangular attention, and then into the large language model (LLM). The LLM retrieves information from the encoded KB to generate an answer.
March 18, 2025 at 10:11 PM
teikan.bsky.social
DeepSeek、こんな地味なのもやってたのか
www.tomshardware.com/pc-component...
DeepSeek brings disruption to AI-optimized parallel file systems, releases powerful new open-source Fire-Flyer File System
3FS brings a new paradigm for AI-HPC training servers, prioritizing random reads above all else
www.tomshardware.com
March 1, 2025 at 11:26 PM
teikan.bsky.social
SLC/QLCをzoneに見せるのはありそうであまりない
ZnH2: Augmenting ZNS-based Storage System with Host-Managed Heterogeneous Zones
yingjia-wang.github.io/papers/znh2-...
名前が凝ってる…
February 26, 2025 at 2:18 PM
teikan.bsky.social
NetAppはちょっとほしい
ascii.jp/elem/000/004...
今年も出るぞ！「手のひらネットワーク機器」第3弾はF5・NEC・NetApp・シュナイダーが監修
累計販売数25万個以上を超える“ITインフラ機器のカプセルトイ”の新作が、今年も登場する。ターリン・インターナショナルは、「手のひらネットワーク機器」シリーズの第3弾を、2025年6月中旬より全国のカプセルトイ自販機コーナーで発売する。
ascii.jp
February 26, 2025 at 1:29 AM
teikan.bsky.social
ASPLOS Proceedings Volume 1
dl.acm.org/doi/proceedi...
Proceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 1 | ACM Conferences
dl.acm.org
February 7, 2025 at 8:13 AM
teikan.bsky.social
EuroSys 2025 Accepted Papers公開
2025.eurosys.org/accepted-pap...
併催のASPLOSはまだ
EuroSys 2025
2025.eurosys.org
January 31, 2025 at 12:56 AM
teikan.bsky.social
dl.acm.org/doi/full/10....
>Compared with other hyperscale infrastructures, we more frequently employ local SSDs to reduce costs, despite the management complexities involved. For instance, imbalanced load distribution can lead to underutilization and stranding of local SSDs.
Meta’s Hyperscale Infrastructure: Overview and Insights | Communications of the ACM
dl.acm.org
January 23, 2025 at 9:44 PM
teikan.bsky.social
持ち上げられる容量で勝負するならM.2とかの方が有利?
spectrum.ieee.org/huge-ssds
AI Is Supersizing Memory Drives
To feed AI's insatiable data appetite, memory drives are getting larger, with no end in sight. Amazingly, this new crop of solid-state drives has the same physical footprint and power budget as their ...
spectrum.ieee.org
January 22, 2025 at 1:21 PM
teikan.bsky.social
😢>The various efforts to introduce host-guided data placement present a sorry picture, we see a cycle of new interfaces being introduced, demonstrated, and deprecated within a couple of years.
arxiv.org/abs/2501.00977
Host-guided data placement: whose job is it anyway?
The increasing demand for SSDs coupled with scaling difficulties have left manufacturers scrambling for newer SSD interfaces which promise better performance and durability. While these interfaces red...
arxiv.org
January 5, 2025 at 11:16 PM
teikan.bsky.social
この論文見落としていた。Azureでの古いパーツの再利用。
ieeexplore.ieee.org/document/106...
・DDR4をCXLで再利用。Pondの手法でがんばれば98%のアプリで5%以下の性能低下で済む
・SSDは7年使ってもほとんどが半分以上寿命残っている。ストライピングすれば性能問題なし
・コスパのいいサーバと比べるとTCOは5%上がるが、炭素排出量が28%下がる
IEEE Spectrum記事もあり。
spectrum.ieee.org/server-reuse
(ウィキペディア図書館でIEEE Xplore見られるようになって便利)
Designing Cloud Servers for Lower Carbon
To mitigate climate change, we must reduce carbon emissions from hyperscale cloud computing. We find that cloud compute servers cause the majority of emissions in a general-purpose cloud. Thus, we mot...
ieeexplore.ieee.org
January 2, 2025 at 10:39 PM
Reposted by YOSHIDA Hideki
phoronix-poster.bsky.social
IBM Deprecating Linux Drivers For CXL Coherent Accelerators & CAPI Flash - https://www.phoronix.com/news/IBM-Deprecates-CXL-CAPI-Flash
IBM Deprecating Linux Drivers For CXL Coherent Accelerators & CAPI Flash
IBM engineers announced they are deprecating the upstream CXL & CXLFLASH Linux kernel drivers with plans to then remove the drivers from the mainline kernel the following cycle. Before getting too worked up when seeing the "cxl: Deprecate driver" patch, this is about the Coherent Accelerator Processor Interface (CAPI) that predates the Compute Express Link...
www.phoronix.com
December 10, 2024 at 11:27 AM