Category: Hardware

Amazon EC2 的 P2 instance

Amazon EC2 為了 GPU 而推出的 P2 type：「New P2 Instance Type for Amazon EC2 – Up to 16 GPUs」。

p2.large 有這樣的規格：

This new instance type incorporates up to 8 NVIDIA Tesla K80 Accelerators, each running a pair of NVIDIA GK210 GPUs. Each GPU provides 12 GB of memory (accessible via 240 GB/second of memory bandwidth), and 2,496 parallel processing cores.

而最大台的 p2.16xlarge 也就是 16 倍... 每小時單價也刷新了之前 x1.32xlarge 的記錄 $13.338/hr (us-east-1)，來到了 $14.4/hr...

另外也推出了 deep learning AMI，內裝了一堆常見支援 GPU 的 ML framework：

In order to help you to make great use of one or more P2 instances, we are launching a Deep Learning AMI today.

Amazon EC2 推出 m4.16xlarge

Amazon EC2 的 m4 系列推出更大台的機器，m4.16xlarge：「Expanding the M4 Instance Type – New M4.16xlarge」。

先前 m4 最大的是是 m4.10xlarge，現在則是往上拉... 比較特別的是這款機器可以吃到 20Gbps (跟之前出的 x1.32xlarge 相同)，更早推出的 m4.10xlarge，以及 {c4,c3,g2,r3,i2,d2}.8xlarge 都只能吃到 10Gbps。

選擇會更活一點... 不過到這個等級的應用應該都規劃過 scale 架構才對？雖然也是有可能有那種買來的 application 啦...

Linode 東京第二機房的訪談 (a.k.a. PR 稿)

前幾天提到的「Linode 東京機房擴充的工作...」，這幾天 Linode 就開始發 PR 稿造勢了：「Behind The Scenes: Details About Upcoming Tokyo 2 DC Launch」，把重點拉出來。

新機房在品川 (Shinagawa)，原文應該是打錯了 (n 跟 b 差一格而已)：

Linode will be opening a new datacenter in Shibagawa ward, Tokyo, Japan, this fall and I was able to interview Linode’s datacenter operations manager, Brett Kaplan, who answered questions I asked regarding the upcoming Tokyo datacenter launch.

目標在 2016Q4 上線：

Soh: Can you say when this new datacenter is expected to come online?
Brett: We are hoping to launch later in Q4 this year.

用的是 Equinix 的機房：(然後這邊的 Shinagawa 是對的 XD)

Soh: So, where did you decide to establish the second Tokyo location?
Brett: We are utilizing an Equinix datacenter in Shinagawa ward Tokyo, Japan.

這樣看起來應該是「Tokyo TY2 Network Interconnection」這個...

Firefox 的大量寫入對 SSD 的影響

在「Firefox is eating your SSD - here is how to fix it」這邊講到 Firefox 寫入對 SSD 的影響，先引用文章裡的解法：

After some digging, I found out that this behavior is controlled by a parameter that you can access through typing “about:config” in the address bar. This parameter is called: —browser.sessionstore.interval

預設是 15 秒一次，作者改成 30 分鐘一次，因此下降了大約 5 倍 (應該可以解讀成 1/6？)：

It is set to 15 seconds by default. In my case, I reset it to a more sane (at least for me) 30 minutes. Since then, I’m only seeing about 2GB written to disk when my workstation is left idle, which still feels like a lot but is 5 times less than before.

據文章後面的 update 說明，Google Chrome 也有類似的情況，不過暫時沒給解法...

PostgreSQL 上，直接將 SSD 的內容送到 GPU 上，加速讀取速度

在 PostgreSQL 上針對讀取檔案到 GPU 上的成果：「GpuScan + SSD-to-GPU Direct DMA」(日文版)、「(EN) GpuScan + SSD-to-GPU Direct DMA」(英文版)。

主要的原因在於雖然已經有 PGStorm 讓 PostgreSQL 把運算丟到 GPU 上加速，但從 disk 讀到 GPU 這段還是有改善的空間：

PG-Strom that is an extension of PostgreSQL to off-load multiple SQL workloads on GPU devices, transparently. It has been developed for four years, and now supports simple scan, tables join, aggregation and projection.
Its prime focus is CPU intensive workloads, on the other hands, it didn't touch storage subsystem of PostgreSQL because the earlier version of PG-Strom assumes all the data set shall be pre-loaded onto physical memory. No need to say, we had a problem when we want to process a data-set larger than physical RAM.

這是成果，可以看到速度快了一倍以上：

這對資料量超過 RAM 大小時的處理會非常有幫助 (因為會有大量的 disk i/o 發生)。

ING Bank 在羅馬尼亞的機房出事...

ING Bank 在羅馬尼亞的機房發生資料損毀：「A Loud Sound Just Shut Down a Bank's Data Center for 10 Hours」。

不過原因是因為火災測試時噴發的音量太大，導致硬碟故障 XDDD

ING Bank’s main data center in Bucharest, Romania, was severely damaged over the weekend during a fire extinguishing test. In what is a very rare but known phenomenon, it was the loud sound of inert gas being released that destroyed dozens of hard drives. The site is currently offline and the bank relies solely on its backup data center, located within a couple of miles’ proximity.

報導給了個測試影片，示範超大的音量會對硬碟有什麼影響：

Lenovo 開始出 Microsoft-only 機了，其他的 OS 都會被鎖住不能用

有人在 Reddit 上抱怨買了一台 Lenovo 的機器卻不能裝 Linux，因為會抓不到硬碟：「Warning: Microsoft Signature PC program now requires that you can't run Linux. Lenovo's recent Ultrabooks among affected systems.」，官方回覆是：

This system has a Signature Edition of Windows 10 Home installed. It is locked per our agreement with Microsoft.

取自這邊：

喔耶...

CPU 指令的速度

在「Infographics: Operation Costs in CPU Clock Cycles」這邊看到張有趣的圖片：

文章大致說明了底層指令速度差異的由來，另外也提到了 C/C++ Function Calls、Indirect and Virtual Calls、Allocations、Kernel Calls、C++ Exceptions 以及 Thread Context Switches 這些高階面向的 CPU clock cycle 差異...

不過重點還是在這張圖 XD

不打開書直接掃描內容

MIT Media Lab 弄出個好玩的東西，可以不打開書直接掃描書的內容：「Can computers read through a book page by page without opening it?」，主標題是「Terahertz time-gated spectral imaging for content extraction through layered structures」。

用 100Ghz 到 3Thz 的電磁波掃描：

In our new study we explore a range of frequencies from 100 Gigahertz to 3 Terahertz (THz) which can penetrate through paper and many other materials.

先前也有類似的方法，用 X-ray 或是超音波，但效果都不好：

Can’t X-ray or ultrasound do this? It may seem that X-ray or ultrasound can also image through a book; however, such techniques lack the contrast of our THz approach for submicron pen or pencil layers compared next to blank paper. These methods have additional drawbacks like cost and ionizing radiation. So while you might be able to hardly detect pages of a closed book if you use a CT scan, you will not be able to see the text. Ultrasound does not have the resolution to detect 20 micron gaps in between the pages of a closed book -distinguishing the ink layers from the blank paper is out of the question for ultrasound. Based on the paper absorption spectrum, we believe that far infrared time resolved systems and THz time domain systems might be the only suitable candidates for investigating paper stacks page by page.

不知道可以進展做到什麼程度，目前只是「能看懂」的程度，品質看起來還是不太夠：