Amazon EC2 的 Trn1 正式開放使用

AWS 自家研發晶片的 trn1.* 上線了:「Amazon EC2 Trn1 Instances for High-Performance Model Training are Now Available」。

先前三家雲端的廠商只有 Google Cloud PlatformTPU 可以 train & evaluate,現在 AWS 推出 AWS Trainium,補上 train 這塊的產品。其中官方宣稱可以比 GPU 架構少 50% 的計算成本:

Trainium-based EC2 Trn1 instances solve this challenge by delivering faster time-to-train while offering up to 50% cost-to-train savings over comparable GPU-based instances.

然後 PyTorchTensorFlow 都有支援:

The Neuron plugin natively integrates with popular ML frameworks, such as PyTorch and TensorFlow.

另外用 neuron-ls 可以看到 Neuron 裝置的資訊,不過沒看懂為什麼要 mask 掉 private ip 的資訊:

大型的 cluster 會使用 Amazon FSx for Lustre 整合提供服務:

For large-scale model training, Trn1 instances integrate with Amazon FSx for Lustre high-performance storage and are deployed in EC2 UltraClusters. EC2 UltraClusters are hyperscale clusters interconnected with a non-blocking petabit-scale network.

但第一波開放的區域有點少,只有萬年美東一區 us-east-1 與美西二區 us-west-2

You can launch Trn1 instances today in the AWS US East (N. Virginia) and US West (Oregon) Regions as On-Demand, Reserved, and Spot Instances or as part of a Savings Plan.

us-east-1trn1.2xlarge 的價錢是 US$1.34375/hr,但沒有實際跑過比較好像沒辦法評估到底行不行...

但總算是擺出個產品對打看看,畢竟要夠大才能去訂製這些東西。

AWS 東京區有 12TB 記憶體的機器了

月初 AWS 宣佈東京區有 u-12tb1.112xlarge 可以用了:「Amazon EC2 High Memory instances with 3, 6, 9, and 12TiB of memory are now available in Asia Pacific (Tokyo) region」。查了一下 on-demend 的價錢是 $131.733/hr,如果一個月以 720 小時來算,要 $94847.76/mo...

沒記錯的話,這種機器應該是要另外申請 limit 才能開,沒辦法說測就測。另外在公告裡面有提到 savings plan ,但沒提到 RI (reserved instance),不確定是不是還沒開 RI 讓使用者買 (不過我記得 savings plan 好像也有類似的折扣結構):

Starting today, Amazon EC2 High Memory instances with 3TiB (u-3tb1.56xlarge), 6TiB (u-6tb1.56xlarge, u-6tb1.112xlarge), 9TiB (u-9tb1.112xlarge), and 12TiB of memory (u-12tb1.112xlarge) are available in Asia Pacific (Tokyo) region. Customers can start using these new High Memory instances with On Demand and Savings Plan purchase options.

這種機器是用暴力解決問題的機器...

EC2 的大記憶體機器又推新規格了...

這次 Amazon EC2 的機器又推出一些新規格了:「Amazon EC2 High Memory Instances now available for on-demand usage」。

然後每次推這些機器的時候都會提到 SAP HANA,都沒有其他的例子可以說... 話說業界就只剩下這套系統是完全都沒在考慮分散式架構嗎 XDDD (完全沒用過)

SAP customers continue to leverage AWS as their platform of choice and innovation. Some are in the early stages of their SAP cloud journeys and are focused on executing their migration. Others have hardened their SAP systems on AWS and are innovating around their core business processes with advanced AWS services.

另外他有提到 24TB 的機器,在 Amazon EC2 Instance Types 這邊可以翻到 u-24tb1.metal

In 2018, we released High Memory Instances in response, which now offer up to 24TB of memory in a single instance.

不過你會發現在 Amazon EC2 On-Demand Pricing 這邊翻不到 24TB 的價錢,先前在「EC2 推出 18TB 與 24TB 的機器...」這邊有過這些機器買三年 RI 才能用,所以這次推出來 12TB 的機器算是隨時租用的機器裡面記憶體最多的了...

u-12tb1.112xlargeus-east-1 的價錢是 USD$109.2/hour,想要玩的人可以測試看看,至少應該玩的動 XD

Google 新推出的 Lyra audio codec

Hacker News Daily 上看到「Lyra audio codec enables high-quality voice calls at 3 kbps bitrate」,講 Google 新推出的 Lyra audio codec:「Lyra: A New Very Low-Bitrate Codec for Speech Compression」,論文在「Generative Speech Coding with Predictive Variance Regularization」這邊可以抓到。

目前 Google 提出來的想法是想辦法在 56kbps 的頻寬下實現還堪用的視訊通話:

Pairing Lyra with new video compression technologies, like AV1, will allow video chats to take place, even for users connecting to the internet via a 56kbps dial-in modem.

這次的突破在於可以使用 3kbps 的頻寬傳輸,但清晰度比 Opus 的 6kbps 效果還好不少。

Google 在文章裡面給了兩個 sample,一個是乾淨背景音,另外一個是吵雜的背景音,跟 Opus 與 Speex 比起來都好很多。

論文是說不需要太高的運算力,但沒翻到 GitHub 之類的 source code,先當作參考:

We provide extensive subjective performance evaluations that show that our system based on generative modeling provides state-of-the-art coding performance at 3 kb/s for real-world speech signals at reasonable computational complexity.

Eventbrite 的 MySQL 升級計畫

在 2021 年看到 EventbiteMySQL 升級計畫:「MySQL High Availability at Eventbrite」。

看起來是 2019 年年初的時候 MySQL 5.1 出問題,後續決定安排升級,在 2019 年年中把系統升級到 MySQL 5.7 (Percona Server 版本):

Our first major hurdle was to get current with our version of MySQL. In July, 2019 we completed the MySQL 5.1 to MySQL 5.7 (v5.7.19-17-log Percona Server to be precise) upgrade across all MySQL instances.

然後看起來是直接在 EC2 上跑,不過這邊提到的空間問題就不太確定了,是真的把 EBS 的空間上限用完嗎?比較常使用的 gp2gp3 上限都是 16TB,不確定是不是真的用到接近爆掉了:

Not only was support for MySQL 5.1 at End-of-Life (more than 5 years ago) but our MySQL 5.1 instances on EC2/AWS had limited storage and we were scheduled to run out of space at the end of July. Our backs were up against the wall and we had to deliver!

另外在升級到 5.7 的時候,順便把本來是 INT 的 primary key 都換成 BIGINT

As part of the cut-over to MySQL 5.7, we also took the opportunity to bake in a number of improvements. We converted all primary key columns from INT to BIGINT to prevent hitting MAX value.

然後系統因為舊版的 Django 沒辦法配合 MySQL 5.7,得升級到 Django 1.6 (要注意 Django 1 系列的最新版是 1.11,看起來光是升級到 1.6 勉強會動就升不上去了?):

In parallel with the MySQL 5.7 upgrade we also Upgraded Django to 1.6 due a behavioral change in MySQL 5.7 related to how transactions/commits were handled for SELECT statements. This behavior change was resulting in errors with older version of Python/Django running on MySQL 5.7

然後採用了 GitHub 家研發的 gh-ost 當作改變 schema 的工具:

In December 2019, the Eventbrite DBRE successfully implemented a table ALTER via gh-ost on one of our larger MySQL tables.

看起來主要的原因是有遇到 pt-online-schema-change 的限制 (在「GitHub 發展出來的 ALTER TABLE 方式」這邊有提到):

Eventbrite had traditionally used pt-online-schema-change (pt-osc) to ALTER MySQL tables in production. pt-osc uses MySQL triggers to move data from the original to the “duplicate” table which is a very expensive operation and can cause replication lag. Matter of fact, it had directly resulted in several outages in H1 of 2019 due to replication lag or breakage.

另外一個引入的技術是 Orchestrator,看起來是先跟 HAProxy 搭配,不過他們打算要再換到 ProxySQL

Next on the list was implementing improvements to MySQL high availability and automatic failover using Orchestrator. In February of 2020 we implemented a new HAProxy layer in front of all DB clusters and we released Orchestrator to production!

Orchestrator can successfully detect the primary failure and promote a new primary. The goal was to implement Orchestrator with HAProxy first and then eventually move to Orchestrator with ProxySQL.

然後最後題到了 Square 研發的 Shift,把 gh-ost 包裝起來變成有個 web UI 可以操作:

2021 還可以看到這類文章還蠻有趣的...

EBS io1 推出可以同時掛到多台的選項

EBS 的 io1 推出了可以同時掛到 16 台 EC2 instance 的選項:「New – Multi-Attach for Provisioned IOPS (io1) Amazon EBS Volumes」。

先看支援的區域,傳統主力區域 (us-east-1 與 eu-west-1) 都支援了,而亞洲區這邊反倒是南韓先支援了:

Multi-Attach for Provisioned IOPS (io1) volumes on Amazon Elastic Block Store (EBS) is available today at no extra charge to customers in the US East (N. Virginia & Ohio), US West (Oregon), EU (Ireland), and Asia Pacific (Seoul) regions.

其中常用的目的是 HA:

Multi-Attach capability makes it easier to achieve higher availability for applications that provide write ordering to maintain storage consistency.

Heartbeat 類的應用應該可以用上這個東西,不過本來就可以透過 command line API 做到 detach & attach,用這個只是少了一個動作...

第二個想到的是,在實體機房的環境下,有些 filesystem (在「Shared-disk file systems」裡面可以翻到一些) 可以同時掛同一個 block storage (通常是透過 SAN),現在在 AWS 上面也可以這樣搞了。

不過 io1 記得不便宜啊...

Amazon RDS 推出了 Connection Pool 的產品

Amazon RDS 推出了 Connection Pool 的產品,叫做 Amazon RDS Proxy:「Introducing Amazon RDS Proxy (Preview)」。

目前支援 MySQL (包括了傳統的與 Aurora 版本的):

Amazon RDS Proxy supports Amazon RDS for MySQL and Amazon Aurora with MySQL compatibility, with support for additional RDS database engines coming soon.

定價策略看起來是依照後端資料庫的 vCPU 計算:

Pricing is simple and predictable: you pay per vCPU of the database instance for which the proxy is enabled.

翻了一下價錢頁是 USD$0.015/vCPU (用 us-east-1 的資料),而如果是 t2 系列的機器,最低是以 2 vCPUs 計算,不是照使用比例算:

RDS Proxy pricing correlates to the number of vCPUs of the database instance for which it is enabled, with a minimum charge for 2 vCPUs.

這樣一個 vCPU 一個月大約要 USD$21.6,算起來頗貴的... 如果 SLA 允許的話,用基本的方式 failover 也許就 ok 了...

如果 SLA 真的要追求到這麼高的話,可以在這些區域測試:

Amazon RDS Proxy is available in preview for RDS MySQL and Aurora MySQL in US East (N. Virginia), US East (Ohio), US West (Oregon), EU West (Ireland), and Asia Pacific (Tokyo) regions. Support for RDS PostgreSQL and Aurora PostgreSQL is coming soon.

Canonical 推出的 Dqlite (High-Availability SQLite)

第一眼看到的時候直接有種不知道 Canonical 在幹什麼的感覺,翻完說明後大概知道可以用的地方,但還是覺得範圍有點小:「Dqlite - High-Availability SQLite」。

一種使用情境是,在 embedded system 上面同步資料的一種方案... 吧?例如網路連線的頻寬或是品質受限,無法順利傳到 Internet 端的伺服器上,所以希望在本地端就可以解決一些事情,但又不方便在本地端直接弄個 PostgreSQL 出來?

Dqlite is a fast, embedded, persistent SQL database with Raft consensus that is perfect for fault-tolerant IoT and Edge devices.

另外一個是用到了 C 實做的 Raft 協定:

Dqlite (“distributed SQLite”) extends SQLite across a cluster of machines, with automatic failover and high-availability to keep your application running. It uses C-Raft, an optimised Raft implementation in C, to gain high-performance transactional consensus and fault tolerance while preserving SQlite’s outstanding efficiency and tiny footprint.

讓 IoT 裝置參與 Raft 嗎... 好像只能說有趣... XD

加快 ls 的速度

看到「When setting an environment variable gives you a 40x speedup」這篇在講 ls 的速度。

文章是由 StanfordSherlock 發出來的,不過看起來跟電視劇沒關係,從網站上的標語「The HPC cluster for all your computing needs」可以看出是 HPC 相關的單位。

在 HPC 環境裡面可以預期單一目錄裡會有很多檔案,所以使用者跑來抱怨 ls 的速度就不算太意外了。不過這次使用者有提到在他自己的 laptop 上跑 ls 反而很快:

It all started from a support question, from a user reporting a usability problem with ls taking several minutes to list the contents of a 15,000+ entries directory on $SCRATCH.

Having thousands of files in a single directory is usually not very file system-friendly, and definitely not recommended. The user knew this already and admitted that wasn’t great, but when he mentioned his laptop was 1,000x faster than Sherlock to list this directory’s contents, of course, it stung. So we looked deeper.

直接跳到後面的結論... 原因是出自於因為需要顯示不同顏色,而需要透過 lstat() 查詢額外的檔案性質 (可執行、setuid 以及 setgid 這些資料),導致速度變慢:

From 13s with the default settings, to 0.3s with a small LS_COLORS tweak, that’s a 40x speedup right there, for the cheap price of not having setuid/setgid or executable files colorized differently.

Of course, this is now setup on Sherlock, for every user’s benefit.

透過設定 LS_COLORS='ex=00:su=00:sg=00:ca=00:',可以讓 lstat() 消失,所以被放進 Sherlock 的預設值了... 而沒有遇到這個問題的環境 (像是有設計好對應的目錄結構),或是想要維持原來的樣子的人,則可以 unset 掉這個值讓輸出還是有色彩差異 :o

JPMorgan Chase 的 WePay 用的 MySQL 架構

看到「Highly Available MySQL Clusters at WePay」這篇講 WePayMySQL 的設計,本來以為是 WeChat 的服務,仔細看查了之後發現原來是 JPMorgan Chase 的服務...

架構在 GCP 上面,本來的 MySQL 是使用 MHA + HAProxy (patch 過的版本,允許動態改變 pool),然後用 Routes 處理 HAProxy 的 failover。

他們遇到的問題是 crash failover 需要至少 30 分鐘的切換時間,另外就是在 GCP 上面跨區時會有的 network partition 問題...

後續架構變得更複雜,讓人懷疑真的有解決問題嗎 XDDD

改用 GitHub 推出的 Orchestrator 架構,然後用兩層 HAProxy 導流 (一層放在 client side,另外一層是原來架構裡面的 load balancer),在加上用 Consul 更新 HAProxy 的資訊?

思考為什麼會有這樣設計 (考慮到金融體系的背景),其實還蠻有趣的...