phpcms 外贸网站模板,做网站的IDE,爱建站小程序功能介绍,wordpress 情侣博客DeepSeek昨天开源了3FS分布式文件系统, 通过180个存储节点提供了 6.6TiB/s的存储性能, 全面支持大模型的训练和推理的KVCache转存以及向量数据库等能力, 每个客户端节点支持40GB/s峰值吞吐用于KVCache查找.
发布后, 我们在阿里云ECS上进行了快速的复现, 并进行了性能测试, ECS…DeepSeek昨天开源了3FS分布式文件系统, 通过180个存储节点提供了 6.6TiB/s的存储性能, 全面支持大模型的训练和推理的KVCache转存以及向量数据库等能力, 每个客户端节点支持40GB/s峰值吞吐用于KVCache查找.
发布后, 我们在阿里云ECS上进行了快速的复现, 并进行了性能测试, ECS在第八代实例中全地域全可用区部署了高性能的eRDMA通信能力, 解决了RDMA超大规模组网的问题, 并且ECS可选的存储介质有 ESSD、EED、本地盘等多种类型.
值得一提的是, 在RDMA大规模组网时通常需要设计基于多路径转发的拥塞控制协议, 例如AWS SRD和UEC, 但是这些协议为了应对多路径转发时的乱序处理, 均不支持标准的RDMA Reliable Connection传输, 因此在适配3FS时会有大量的工作, 而eRDMA实现了高性能多路径转发及拥塞控制,并且完全兼容标准RDMA Reliable Connection传输, 无需修改任何3FS代码就可以直接运行.
为了对标3FS的官方部署指南, 我们这次测试中选择了本地盘实例构建了5个存储节点, 并且通过5个client进行了测试, 经过测试所有节点都能够打满实例产品规格的带宽(单机100Gbps). ECS 9代实例将普遍标配400Gbps CIPU 2.0, 带宽和DeepSeek 3FS线下部署规格一致, 后续可以根据用户的需求提供新的实例规格满足业务需求. 对于缺少RDMA和相关存储测试环境的研究者和开源生态的贡献者, 可以通过如下文档在阿里云上基于eRDMA构建3FS并进行后续的测试和开发. 后续我们将对3fs进行更多的分析. 本文结构如下
1. DeepSeek 3FS分布式存储概述
2. 安装和编译3fs
2.1 构建编译环境
2.2 编译3fs
2.3 制作镜像
3. 部署3FS
3.1 安装ClickHouse和FoundationDB
3.2 配置监控服务
3.3 配置Admin Client
3.4 配置Mgmtd Service
3.5 配置Meta Service
3.6 配置Storage Service
3.7 配置3FS
3.8 配置FUSE Client
4. 性能测试1. DeepSeek 3FS分布式存储概述
对3FS的关注大概是在2019年幻方有一篇文章介绍3FS时就在关注它, 当时只有一个record格式的git[1]. 作为量化交易的同行, 我在2014年搭建自己的私募量化平台时也在做一些分布式内存数据库的实现, 主要用途就是模型需要快速的从大量tick数据里抓取数据, 另外一些回测框架也需要极高的I/O处理能力, 类似于今天开源的另一个小项目smallpond[2]:
df sp.partial_sql(SELECT ticker, min(price), max(price) FROM {0} GROUP BY ticker, df)
df.write_parquet(output/)
print(df.to_pandas())正是这些原来在幻方量化使用的高性能分布式文件系统, 这一次用在了DeepSeek大模型的训练和推理上.下图展示了在 3FS 集群上进行的读压力测试的吞吐量。该集群由 180 个存储节点组成每个节点配备 2×200Gbps InfiniBand 网卡和十六块 14TiB NVMe SSD。大约 500个客户端节点被用于读压力测试每个客户端节点配置了 1x200Gbps InfiniBand 网卡。最终的累计读吞吐量达到约 6.6 TiB/s包括来自训练作业的背景流量。 它采用了基于CRAQ的链式replication机制实现了数据的强一致性, 使应用能够以本地无关的方式访问分布在数百个服务器上的数千个SSD的存储资源. 然后还集成了用于LLM推理优化的KVCache服务, 客户端的峰值吞吐高达40GB/s 详细的3FS架构和实现分析我们将在后续的文章中进行分析, 这一篇主要讲解如何基于云服务和eRDMA安装部署并进行性能测试.
2. 安装和编译3fs
首先我们在阿里云上申请一个ecs.g8i.4xlarge实例作为编译环境使用. 注意在创建实例的时候,选择unbuntu 22.04, 并勾选eRDMA驱动安装 同时在弹性网卡中勾选弹性RDMA接口 2.1 构建编译环境
首先安装编译需要的package
# for Ubuntu 22.04.
apt install cmake libuv1-dev liblz4-dev liblzma-dev libdouble-conversion-dev libprocps-dev libdwarf-dev libunwind-dev \libaio-dev libgflags-dev libgoogle-glog-dev libgtest-dev libgmock-dev clang-format-14 clang-14 clang-tidy-14 lld-14 \libgoogle-perftools-dev google-perftools libssl-dev ccache gcc-12 g-12 libboost-all-dev然后安装libfuse, 需要注意使用fuse3.16以上的版本
wget https://github.com/libfuse/libfuse/releases/download/fuse-3.16.1/fuse-3.16.1.tar.gz
tar vzxf fuse-3.16.1.tar.gz
cd fuse-3.16.1/
mkdir build; cd build
apt install meson
meson setup ..
ninja ; ninja install安装rust工具链
#rust toolchains
curl --proto https --tlsv1.2 -sSf https://sh.rustup.rs | sh安装foundation db
Download at https://github.com/apple/foundationdb/releases/tag/7.3.63wget https://github.com/apple/foundationdb/releases/download/7.3.63/foundationdb-clients_7.3.63-1_amd64.debwget https://github.com/apple/foundationdb/releases/download/7.3.63/foundationdb-server_7.3.63-1_amd64.debdpkg -i foundationdb-clients_7.3.63-1_amd64.deb
dpkg -i foundationdb-server_7.3.63-1_amd64.deb2.2 编译3fs
按照如下方式下载和编译3fs
git clone https://github.com/deepseek-ai/3fs
cd 3fs
git submodule update --init --recursive
./patches/apply.sh
cmake -S . -B build -DCMAKE_CXX_COMPILERclang-14 -DCMAKE_C_COMPILERclang-14 -DCMAKE_BUILD_TYPERelWithDebInfo -DCMAKE_EXPORT_COMPILE_COMMANDSON
cmake --build build -j 32检查编译输出的binary
root3fs-1:~/3fs# ls -lrt build/bin/
total 2289660
-rwxr-xr-x 1 root root 148429888 Feb 28 17:41 hf3fs-admin
-rwxr-xr-x 1 root root 105192704 Feb 28 17:41 monitor_collector_main
-rwxr-xr-x 1 root root 178870448 Feb 28 17:42 mgmtd_main
-rwxr-xr-x 1 root root 172303184 Feb 28 17:47 migration_main
-rwxr-xr-x 1 root root 363821208 Feb 28 17:47 admin_cli
-rwxr-xr-x 1 root root 174688488 Feb 28 17:47 simple_example_main
-rwxr-xr-x 1 root root 284427704 Feb 28 17:48 meta_main
-rwxr-xr-x 1 root root 395983072 Feb 28 17:48 storage_bench
-rwxr-xr-x 1 root root 311693016 Feb 28 17:48 storage_main
-rwxr-xr-x 1 root root 209211768 Feb 28 17:48 hf3fs_fuse_main2.3 制作镜像
编译完成后对该台机器制作一个镜像用于后续部署 3. 部署3FS
我们采用官方文档推荐的方式, 在阿里云上创建1个ecs.g8i.48xlarge作为meta服务器和5个ecs.i4.32xlarge实例作为存储服务器. Node 实例类型 IP Memory SSD RDMA meta ecs.g8i.48xlarge 10.99.0.1 256GB - eRDMA storage1 ecs.i4.32xlarge 10.99.0.2 1024GB 4TB × 8 eRDMA storage2 ecs.i4.32xlarge 10.99.0.3 1024GB 4TB × 8 eRDMA storage3 ecs.i4.32xlarge 10.99.0.4 1024GB 4TB × 8 eRDMA storage4 ecs.i4.32xlarge 10.99.0.5 1024GB 4TB × 8 eRDMA storage5 ecs.i4.32xlarge 10.99.0.6 1024GB 4TB × 8 eRDMA fuseclient1 ecs.ebmg8i.48xlarge 10.99.0.101 1024GB - eRDMA fuseclient2 ecs.ebmg8i.48xlarge 10.99.0.102 1024GB - eRDMA fuseclient3 ecs.ebmg8i.48xlarge 10.99.0.103 1024GB - eRDMA fuseclient4 ecs.ebmg8i.48xlarge 10.99.0.104 1024GB - eRDMA fuseclient5 ecs.ebmg8i.48xlarge 10.99.0.105 1024GB - eRDMA
启动后, 将所有的eRDMA模式调成compatmode1
rmmod erdma
modprobe erdma compat_mode1修改配置文件中的max_sge
cd ~/3fs/configs
sed -i s/max_sge 16/max_sge 1/g grep -rl max_sge另外由于3FS使用了mellanox网卡的ibdev2netdev,在执行3fs命令时会调用, 因此我们在eRDMA环境,我们需要构造一个命令输出.采用如下方式添加脚本
vim /usr/sbin/ibdev2netdev
添加如下内容#!/bin/bash
echo erdma_0 port 1 eth0 (Up)
保存退出后, 修改为可执行
chmod x /usr/sbin/ibdev2netdev然后将meta对应的ip填入每个节点的/etc/hosts
vim /etc/hosts
#添加
10.99.0.1 meta每个节点的服务和相应的配置文件和官方建议相同,如下所示: Service Binary Config files NodeID Node monitor monitor_collector_main monitor_collector_main.toml - meta admin_cli admin_cli admin_cli.toml fdb.cluster - meta storage1 storage2 storage3 storage4 storage5 mgmtd mgmtd_main mgmtd_main_launcher.toml mgmtd_main.toml mgmtd_main_app.toml fdb.cluster 1 meta meta meta_main meta_main_launcher.toml meta_main.toml meta_main_app.toml fdb.cluster 100 meta storage storage_main storage_main_launcher.toml storage_main.toml storage_main_app.toml 10001~10005 storage1 storage2 storage3 storage4 storage5 client hf3fs_fuse_main hf3fs_fuse_main_launcher.toml hf3fs_fuse_main.toml - meta
3.1 安装ClickHouse和FoundationDB
由于复用了编译环境的镜像已经安装了FoundationDB, 因此仅需在meta节点安装ClickHouse
安装clickhouse, 可以参考https://clickhouse.com/docs/install
sudo apt-get install -y apt-transport-https ca-certificates curl gnupg
curl -fsSL https://packages.clickhouse.com/rpm/lts/repodata/repomd.xml.key | sudo gpg --dearmor -o /usr/share/keyrings/clickhouse-keyring.gpgARCH$(dpkg --print-architecture)
echo deb [signed-by/usr/share/keyrings/clickhouse-keyring.gpg arch${ARCH}] https://packages.clickhouse.com/deb stable main | sudo tee /etc/apt/sources.list.d/clickhouse.list
sudo apt-get updatesudo apt-get install -y clickhouse-server clickhouse-client
#在安装的时候会要求输入密码, 此时我们统一输入eRDMA123!!使用如下方式开启clickhouse服务 sudo clickhouse start
然后使用安装时的密码验证登陆
root3fs-meta:~# clickhouse-client --password eRDMA123!!
ClickHouse client version 25.2.1.3085 (official build).
Connecting to localhost:9000 as user default.
Connected to ClickHouse server version 25.2.1.Warnings:* Delay accounting is not enabled, OSIOWaitMicroseconds will not be gathered. You can enable it using echo 1 /proc/sys/kernel/task_delayacct or by using sysctl.3fs-meta :)然后退出, 并采用如下命令创建Metric table
clickhouse-client --password eRDMA123!! -n ~/3fs/deploy/sql/3fs-monitor.sql3.2 配置监控服务
仅在meta节点配置安装monitor_collector服务.
mkdir -p /opt/3fs/{bin,etc}
mkdir -p /var/log/3fs
cp ~/3fs/build/bin/monitor_collector_main /opt/3fs/bin
cp ~/3fs/configs/monitor_collector_main.toml /opt/3fs/etc修改monitor_collector_main.toml如下所示
vim /opt/3fs/etc/monitor_collector_main.toml#最后一段修改为Host IP, 预配置密码, 用户名默认为default, 端口号默认为9000[server.monitor_collector.reporter.clickhouse]
db 3fs
host 127.0.0.1
passwd eRDMA123!!
port 9000
user default启动monitor_collector服务如下
cp ~/3fs/deploy/systemd/monitor_collector_main.service /usr/lib/systemd/system
systemctl start monitor_collector_main检查服务状态
root3fs-meta:/opt/3fs/etc# systemctl status monitor_collector_main
● monitor_collector_main.service - monitor_collector_main ServerLoaded: loaded (/lib/systemd/system/monitor_collector_main.service; disabled; vendor preset: enabled)Active: active (running) since Fri 2025-02-28 21:09:06 CST; 15s agoMain PID: 14401 (monitor_collect)Tasks: 58 (limit: 629145)Memory: 258.4MCPU: 113msCGroup: /system.slice/monitor_collector_main.service└─14401 /opt/3fs/bin/monitor_collector_main --cfg /opt/3fs/etc/monitor_collector_main.toml3.3 配置Admin Client
在所有节点安装admin_cli
mkdir -p /opt/3fs/{bin,etc}
rsync -avz meta:~/3fs/build/bin/admin_cli /opt/3fs/bin
rsync -avz meta:~/3fs/configs/admin_cli.toml /opt/3fs/etc
rsync -avz meta:/etc/foundationdb/fdb.cluster /opt/3fs/etc更新admin_cli.toml
vim /opt/3fs/etc/admin_cli.toml
##更新如下内容cluster_id stage[fdb]
clusterFile /opt/3fs/etc/fdb.clusteradmin_cli的使用帮助文档可以输入
root3fs-meta:/opt/3fs/etc# /opt/3fs/bin/admin_cli -cfg /opt/3fs/etc/admin_cli.toml help
bench Usage: bench [--rank VAR] [--timeout VAR] [--coroutines VAR] [--seconds VAR] [--remove] path
cd Usage: cd [-L] [--inode] path
checksum Usage: checksum [--list] [--batch VAR] [--md5] [--fillZero] [--output VAR] path
create Usage: create [--perm VAR] [--chain-table-id VAR] [--chain-table-ver VAR] [--chain-list VAR] [--chunk-size VAR] [--stripe-size VAR] path
create-range Usage: create-range [--concurrency VAR] prefix inclusive_start exclusive_end
create-target Usage: create-target --node-id VAR --disk-index VAR --target-id VAR --chain-id VAR [--add-chunk-size] [--chunk-size VAR...] [--use-new-chunk-engine]
create-targets Usage: create-targets --node-id VAR [--disk-index VAR...] [--allow-existing-target] [--add-chunk-size] [--use-new-chunk-engine]
current-user Usage: current-user
decode-user-token Usage: decode-user-token token
drop-user-cache Usage: drop-user-cache [--uid VAR] [--all]
dump-chain-table Usage: dump-chain-table [--version VAR] tableId csv-file-path
dump-chains Usage: dump-chains csv-file-path-prefix
dump-chunkmeta Usage: dump-chunkmeta [--chain-ids VAR...] [--chunkmeta-dir VAR] [--parquet-format] [--only-head] [--parallel VAR]
dump-dentries Usage: dump-dentries [--num-dentries-perfile VAR] [--fdb-cluster-file VAR] [--dentry-dir VAR] [--threads VAR]
dump-inodes Usage: dump-inodes [--num-inodes-perfile VAR] [--fdb-cluster-file VAR] [--inode-dir VAR] [--parquet-format] [--all-inodes] [--threads VAR]
.....3.4 配置Mgmtd Service
mgmtd仅在meta节点安装.首先拷贝文件
cp ~/3fs/build/bin/mgmtd_main /opt/3fs/bin
cp ~/3fs/configs/{mgmtd_main.toml,mgmtd_main_launcher.toml,mgmtd_main_app.toml} /opt/3fs/etc修改配置文件, 将mgmtd配置文件mgmtd_main_app.toml定义node_id 1
vim /opt/3fs/etc/mgmtd_main_app.toml
##修改
node_id 1修改/opt/3fs/etc/mgmtd_main_launcher.toml中的cluster_id和clusterFile
cluster_id stage[fdb]
clusterFile /opt/3fs/etc/fdb.cluster修改mgmtd_main.toml将remoteip修改为meta服务器地址
[common.monitor.reporters.monitor_collector]
remote_ip 10.99.0.1:10000配置完成后, 初始化集群
/opt/3fs/bin/admin_cli -cfg /opt/3fs/etc/admin_cli.toml init-cluster --mgmtd /opt/3fs/etc/mgmtd_main.toml 1 1048576 16Init filesystem, root directory layout: chain table ChainTableId(1), chunksize 1048576, stripesize 16Init config for MGMTD version 1其中参数1代表chainTable ID, 1048576代表chunksize, 16代表file strip size.然后启动服务并验证
cp ~/3fs/deploy/systemd/mgmtd_main.service /usr/lib/systemd/system
systemctl start mgmtd_mainroot3fs-meta:/opt/3fs/etc# systemctl status mgmtd_main
● mgmtd_main.service - mgmtd_main ServerLoaded: loaded (/lib/systemd/system/mgmtd_main.service; disabled; vendor preset: enabled)Active: active (running) since Fri 2025-02-28 21:33:46 CST; 27s agoMain PID: 16375 (mgmtd_main)Tasks: 36 (limit: 629145)Memory: 192.7MCPU: 123msCGroup: /system.slice/mgmtd_main.service└─16375 /opt/3fs/bin/mgmtd_main --launcher_cfg /opt/3fs/etc/mgmtd_main_launcher.toml --app-cfg /opt/3fs/etc/mgmtd_main_app.toml然后采用如下命令检查节点:
root3fs-meta:~# /opt/3fs/bin/admin_cli -cfg /opt/3fs/etc/admin_cli.toml --config.mgmtd_client.mgmtd_server_addresses [RDMA://10.99.0.1:8000] list-nodes
Id Type Status Hostname Pid Tags LastHeartbeatTime ConfigVersion ReleaseVersion
1 MGMTD PRIMARY_MGMTD 3fs-meta 17434 [] N/A 1(UPTODATE) 250228-dev-1-999999-824fbf5c3.5 配置Meta Service
该服务仅在meta服务器安装, 拷贝文件如下所示
cp ~/3fs/build/bin/meta_main /opt/3fs/bin
cp ~/3fs/configs/{meta_main_launcher.toml,meta_main.toml,meta_main_app.toml} /opt/3fs/etc修改meta_main_app.toml中的node_id 100. 修改meta_main_launcher.toml中的 cluster_id, clusterFile
cluster_id stage[mgmtd_client]
mgmtd_server_addresses [RDMA://10.99.0.1:8000]修改meta_main.toml如下:
[server.mgmtd_client]
mgmtd_server_addresses [RDMA://10.99.0.1:8000][common.monitor.reporters.monitor_collector]
remote_ip 10.99.0.1:10000[server.fdb]
clusterFile /opt/3fs/etc/fdb.cluster更新配置如下
/opt/3fs/bin/admin_cli -cfg /opt/3fs/etc/admin_cli.toml --config.mgmtd_client.mgmtd_server_addresses [RDMA://10.99.0.1:8000] set-config --type META --file /opt/3fs/etc/meta_main.toml启动服务
cp ~/3fs/deploy/systemd/meta_main.service /usr/lib/systemd/system
systemctl start meta_mainroot3fs-meta:~# systemctl status meta_main
● meta_main.service - meta_main ServerLoaded: loaded (/lib/systemd/system/meta_main.service; disabled; vendor preset: enabled)Active: active (running) since Fri 2025-02-28 22:37:58 CST; 7s agoMain PID: 17709 (meta_main)Tasks: 64 (limit: 629145)Memory: 408.9MCPU: 250msCGroup: /system.slice/meta_main.service└─17709 /opt/3fs/bin/meta_main --launcher_cfg /opt/3fs/etc/meta_main_launcher.toml --app-cfg /opt/3fs/etc/meta_main_app.toml
检查节点
root3fs-meta:~# /opt/3fs/bin/admin_cli -cfg /opt/3fs/etc/admin_cli.toml --config.mgmtd_client.mgmtd_server_addresses [RDMA://10.99.0.1:8000] list-nodes
Id Type Status Hostname Pid Tags LastHeartbeatTime ConfigVersion ReleaseVersion
1 MGMTD PRIMARY_MGMTD 3fs-meta 17434 [] N/A 1(UPTODATE) 250228-dev-1-999999-824fbf5c
100 META HEARTBEAT_CONNECTED 3fs-meta 17709 [] 2025-02-28 22:38:28 1 250228-dev-1-999999-824fbf5c3.6 配置Storage Service
在所有存储节点启用storage服务, 由于我们每个节点只有8块盘, 配置挂载如下:
mkdir -p /storage/data{0..7}
mkdir -p /var/log/3fs
for i in {0..7};do mkfs.xfs -L data${i} /dev/nvme${i}n1;mount -o noatime,nodiratime -L data${i} /storage/data${i};done
mkdir -p /storage/data{0..7}/3fsroot3fs-storage001:~# df -kh | grep nvme
/dev/nvme1n1 3.5T 25G 3.5T 1% /storage/data1
/dev/nvme2n1 3.5T 25G 3.5T 1% /storage/data2
/dev/nvme3n1 3.5T 25G 3.5T 1% /storage/data3
/dev/nvme4n1 3.5T 25G 3.5T 1% /storage/data4
/dev/nvme5n1 3.5T 25G 3.5T 1% /storage/data5
/dev/nvme6n1 3.5T 25G 3.5T 1% /storage/data6
/dev/nvme7n1 3.5T 25G 3.5T 1% /storage/data7
/dev/nvme0n1 3.5T 25G 3.5T 1% /storage/data0增加aio请求的最大数
sysctl -w fs.aio-max-nr67108864修改meta节点的原始配置文件~/3fs/configs/storage_main_launcher.toml中的clusterid和管理地址 vim ~/3fs/configs/storage_main_launcher.tomlcluster_id stage[mgmtd_client]
mgmtd_server_addresses [RDMA://10.99.0.1:8000]修改~/3fs/configs/storage_main.toml中的IP地址和target path
vim ~/3fs/configs/storage_main.toml[server.mgmtd]
mgmtd_server_address [RDMA://10.99.0.1:8000][common.monitor.reporters.monitor_collector]
remote_ip 10.99.0.1:10000[server.targets]
target_paths [/storage/data0/3fs,/storage/data1/3fs,/storage/data2/3fs,/storage/data3/3fs,/storage/data4/3fs,/storage/data5/3fs,/storage/data6/3fs,/storage/data7/3fs]从meta节点拷贝执行文件和配置文件
rsync -avz meta:~/3fs/build/bin/storage_main /opt/3fs/bin
rsync -avz meta:~/3fs/configs/{storage_main_launcher.toml,storage_main.toml,storage_main_app.toml} /opt/3fs/etc每个存储节点修改/opt/3fs/etc/storage_main_app.toml中的node_id, 五台机器分别为10001~10005
然后每个存储节点更新
/opt/3fs/bin/admin_cli -cfg /opt/3fs/etc/admin_cli.toml --config.mgmtd_client.mgmtd_server_addresses [RDMA://10.99.0.1:8000] set-config --type STORAGE --file /opt/3fs/etc/storage_main.toml最后启动并验证服务
rsync -avz meta:~/3fs/deploy/systemd/storage_main.service /usr/lib/systemd/system
systemctl start storage_mainroot3fs-storage001:/opt/3fs/etc# systemctl status storage_main
● storage_main.service - storage_main ServerLoaded: loaded (/lib/systemd/system/storage_main.service; disabled; vendor preset: enabled)Active: active (running) since Fri 2025-02-28 23:02:07 CST; 30s agoMain PID: 7788 (storage_main)Tasks: 242 (limit: 629145)Memory: 9.5GCPU: 10.017sCGroup: /system.slice/storage_main.service
检查系统节点:
root3fs-storage001:~# /opt/3fs/bin/admin_cli -cfg /opt/3fs/etc/admin_cli.toml --config.mgmtd_client.mgmtd_server_addresses [RDMA://10.99.0.1:8000] list-nodes
/root/.profile: line 10: /.cargo/env: No such file or directory
Id Type Status Hostname Pid Tags LastHeartbeatTime ConfigVersion ReleaseVersion
1 MGMTD PRIMARY_MGMTD 3fs-meta 17434 [] N/A 1(UPTODATE) 250228-dev-1-999999-824fbf5c
100 META HEARTBEAT_CONNECTED 3fs-meta 17709 [] 2025-02-28 23:03:19 2(UPTODATE) 250228-dev-1-999999-824fbf5c
10001 STORAGE HEARTBEAT_CONNECTED 3fs-storage001 7788 [] 2025-02-28 23:03:20 5(UPTODATE) 250228-dev-1-999999-824fbf5c
10002 STORAGE HEARTBEAT_CONNECTED 3fs-storage002 9025 [] 2025-02-28 23:03:22 5(UPTODATE) 250228-dev-1-999999-824fbf5c
10003 STORAGE HEARTBEAT_CONNECTED 3fs-storage003 6745 [] 2025-02-28 23:03:20 5(UPTODATE) 250228-dev-1-999999-824fbf5c
10004 STORAGE HEARTBEAT_CONNECTED 3fs-storage004 7309 [] 2025-02-28 23:03:21 5(UPTODATE) 250228-dev-1-999999-824fbf5c
10005 STORAGE HEARTBEAT_CONNECTED 3fs-storage005 6776 [] 2025-02-28 23:03:19 5(UPTODATE) 250228-dev-1-999999-824fbf5c3.7 配置3FS
创建管理员
root3fs-meta:~/3fs/configs# /opt/3fs/bin/admin_cli -cfg /opt/3fs/etc/admin_cli.toml --config.mgmtd_client.mgmtd_server_addresses [RDMA://10.99.0.1:8000] user-add --root --admin 0 root
Uid 0
Name root
Token AADDI7y8QAUtUR2wCeuDI5(Expired at N/A)
IsRootUser true
IsAdmin true
Gid 0
SupplementaryGids将token保存在/opt/3fs/etc/token.txt中.
然后创建chain Table, 首先安装python相关的依赖
pip3 install -r ~/3fs/deploy/data_placement/requirements.txt然后执行data_placement计算命令
root3fs-meta# python3 ~/3fs/deploy/data_placement/src/model/data_placement.py \-ql -relax -type CR --num_nodes 5 --replication_factor 3 --min_targets_per_disk 62025-02-28 23:23:06.821 | INFO | __main__:run:125 - solving model with appsi_highs #0: DataPlacementModel-v5,b10,r6,k3,λ2,lb1,ub0
2025-02-28 23:23:06.821 | INFO | __main__:build_model:182 - self.num_nodes5 self.num_targets_per_disk6 self.group_size3 self.num_groups10 self.qlinearizeTrue self.relax_lb1 self.relax_ub0
2025-02-28 23:23:06.821 | INFO | __main__:build_model:192 - self.sum_recovery_traffic_per_failure6 self.max_recovery_traffic_on_peer2
2025-02-28 23:23:06.821 | INFO | __main__:build_model:196 - self.all_targets_usedTrue self.balanced_peer_trafficFalse
2025-02-28 23:23:06.821 | INFO | __main__:build_model:197 - self.num_targets_used30 self.num_targets_total30
2025-02-28 23:23:06.839 | INFO | __main__:build_model:272 - lower bound imposed on peer traffic: self.relax_lb1 self.qlinearizeTrue self.all_targets_usedTrue
Running HiGHS 1.8.0 (git hash: eda5cbe): Copyright (c) 2024 HiGHS under MIT licence terms1 0 1 100.00% inf inf inf 132 16 7 299 0.0sSolving reportStatus InfeasiblePrimal bound infDual bound infGap infSolution status -Timing 0.02 (total)0.00 (presolve)0.00 (postsolve)Nodes 1LP iterations 299 (total)0 (strong br.)109 (separation)0 (heuristics)Nodes | BB Tree | Objective Bounds | Dynamic Constraints | WorkProc. InQueue | Leaves Expl. | BestBound BestSol Gap | Cuts InLp Confl. | LpIters Time0 0 0 0.00% 1 inf inf 0 0 0 0 0.0s
Objective function is integral with scale 1
Coefficient ranges:Matrix [1e00, 1e00]Cost [0e00, 0e00]Bound [1e00, 1e00]RHS [1e00, 6e00]
Presolving model
335 rows, 150 cols, 1000 nonzeros 0s
325 rows, 150 cols, 900 nonzeros 0sSolving MIP model with:325 rows150 cols (150 binary, 0 integer, 0 implied int., 0 continuous)900 nonzeros0 0 0 0.00% 1 inf inf 0 0 4 190 0.0s
2025-02-28 23:23:06.879 | ERROR | __main__:run:133 - cannot find solution for current params: infeasible:
- Status: errorTermination condition: infeasibleTermination message: TerminationCondition.infeasible2025-02-28 23:23:06.879 | INFO | __main__:run:125 - solving model with appsi_highs #1: DataPlacementModel-v5,b10,r6,k3,λ2,lb1,ub1
2025-02-28 23:23:06.879 | INFO | __main__:build_model:182 - self.num_nodes5 self.num_targets_per_disk6 self.group_size3 self.num_groups10 self.qlinearizeTrue self.relax_lb1 self.relax_ub1
2025-02-28 23:23:06.880 | INFO | __main__:build_model:192 - self.sum_recovery_traffic_per_failure6 self.max_recovery_traffic_on_peer2
2025-02-28 23:23:06.880 | INFO | __main__:build_model:196 - self.all_targets_usedTrue self.balanced_peer_trafficFalse
2025-02-28 23:23:06.880 | INFO | __main__:build_model:197 - self.num_targets_used30 self.num_targets_total30
2025-02-28 23:23:06.882 | INFO | __main__:build_model:272 - lower bound imposed on peer traffic: self.relax_lb1 self.qlinearizeTrue self.all_targets_usedTrue
Running HiGHS 1.8.0 (git hash: eda5cbe): Copyright (c) 2024 HiGHS under MIT licence terms1 0 1 100.00% 1 1 0.00% 57 4 3 194 0.0sSolving reportStatus OptimalPrimal bound 1Dual bound 1Gap 0% (tolerance: 0.01%)Solution status feasible1 (objective)0 (bound viol.)0 (int. viol.)0 (row viol.)0 0 0 0.00% 1 inf inf 0 0 3 181 0.0s0 0 0 0.00% 1 inf inf 0 0 0 0 0.0s
Objective function is integral with scale 1
Coefficient ranges:Matrix [1e00, 1e00]Cost [0e00, 0e00]Bound [1e00, 1e00]RHS [1e00, 6e00]
Presolving model
335 rows, 150 cols, 1000 nonzeros 0s
325 rows, 150 cols, 900 nonzeros 0sSolving MIP model with:325 rows150 cols (150 binary, 0 integer, 0 implied int., 0 continuous)900 nonzerosNodes | BB Tree | Objective Bounds | Dynamic Constraints | WorkProc. InQueue | Leaves Expl. | BestBound BestSol Gap | Cuts InLp Confl. | LpIters TimeTiming 0.01 (total)0.00 (presolve)0.00 (postsolve)Nodes 1LP iterations 194 (total)0 (strong br.)13 (separation)0 (heuristics)
2025-02-28 23:23:06.906 | SUCCESS | __main__:solve:165 - optimal solution:
- Status: okTermination condition: optimalTermination message: TerminationCondition.optimal2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 1,2: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 1,3: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 1,4: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 1,5: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 2,1: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 2,3: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 2,4: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 2,5: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 3,1: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 3,2: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 3,4: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 3,5: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 4,1: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 4,2: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 4,3: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 4,5: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 5,1: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 5,2: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 5,3: 1.5
2025-02-28 23:23:06.907 | DEBUG | __main__:check_solution:322 - 5,4: 1.5
2025-02-28 23:23:06.907 | INFO | __main__:check_solution:331 - min_peer_traffic1.5 max_peer_traffic1.5
2025-02-28 23:23:06.907 | INFO | __main__:check_solution:332 - total_traffic30.0 max_total_traffic30
2025-02-28 23:23:07.068 | SUCCESS | __main__:run:148 - saved solution to: output/DataPlacementModel-v_5-b_10-r_6-k_3-λ_2-lb_1-ub_1
然后执行产生chainTable
python3 ~/3fs/deploy/data_placement/src/setup/gen_chain_table.py \--chain_table_type CR --node_id_begin 10001 --node_id_end 10005 \--num_disks_per_node 8 --num_targets_per_disk 6 \--target_id_prefix 1 --chain_id_prefix 9 \--incidence_matrix_path output/DataPlacementModel-v_5-b_10-r_6-k_3-λ_2-lb_1-ub_1/incidence_matrix.pickle检查output目录是否产生了如下文件
root3fs-meta:/opt/3fs# ls -lrt output-rw-r--r-- 1 root root 808 Feb 28 23:24 generated_chain_table.csv
-rw-r--r-- 1 root root 3955 Feb 28 23:24 generated_chains.csv
-rw-r--r-- 1 root root 27600 Feb 28 23:24 create_target_cmd.txt创建storage target
/opt/3fs/bin/admin_cli --cfg /opt/3fs/etc/admin_cli.toml --config.mgmtd_client.mgmtd_server_addresses [RDMA://10.99.0.1:8000] --config.user_info.token $(/opt/3fs/etc/token.txt) output/create_target_cmd.txt上传chains 和 chain table到mgmtd service
/opt/3fs/bin/admin_cli --cfg /opt/3fs/etc/admin_cli.toml --config.mgmtd_client.mgmtd_server_addresses [RDMA://10.99.0.1:8000] --config.user_info.token $(/opt/3fs/etc/token.txt) upload-chains output/generated_chains.csv/opt/3fs/bin/admin_cli --cfg /opt/3fs/etc/admin_cli.toml --config.mgmtd_client.mgmtd_server_addresses [RDMA://10.99.0.1:8000] --config.user_info.token $(/opt/3fs/etc/token.txt) upload-chain-table --desc stage 1 output/generated_chain_table.csv
检查是否上传成功
# /opt/3fs/bin/admin_cli -cfg /opt/3fs/etc/admin_cli.toml --config.mgmtd_client.mgmtd_server_addresses [RDMA://10.99.0.1:8000] list-chains900800001 1 1 SERVING [] 101000100801(SERVING-UPTODATE) 101000200801(SERVING-UPTODATE) 101000400801(SERVING-UPTODATE)
900800002 1 1 SERVING [] 101000200802(SERVING-UPTODATE) 101000300801(SERVING-UPTODATE) 101000500801(SERVING-UPTODATE)
900800003 1 1 SERVING [] 101000100802(SERVING-UPTODATE) 101000200803(SERVING-UPTODATE) 101000300802(SERVING-UPTODATE)
900800004 1 1 SERVING [] 101000100803(SERVING-UPTODATE) 101000200804(SERVING-UPTODATE) 101000500802(SERVING-UPTODATE)
900800005 1 1 SERVING [] 101000100804(SERVING-UPTODATE) 101000400802(SERVING-UPTODATE) 101000500803(SERVING-UPTODATE)
900800006 1 1 SERVING [] 101000200805(SERVING-UPTODATE) 101000300803(SERVING-UPTODATE) 101000400803(SERVING-UPTODATE)
900800007 1 1 SERVING [] 101000200806(SERVING-UPTODATE) 101000400804(SERVING-UPTODATE) 101000500804(SERVING-UPTODATE)
900800008 1 1 SERVING [] 101000100805(SERVING-UPTODATE) 101000300804(SERVING-UPTODATE) 101000400805(SERVING-UPTODATE)
900800009 1 1 SERVING [] 101000300805(SERVING-UPTODATE) 101000400806(SERVING-UPTODATE) 101000500805(SERVING-UPTODATE)# /opt/3fs/bin/admin_cli -cfg /opt/3fs/etc/admin_cli.toml --config.mgmtd_client.mgmtd_server_addresses [RDMA://10.99.0.1:8000] list-chain-tablesChainTableId ChainTableVersion ChainCount ReplicaCount Desc
1 1 80 3 stage3.8 配置FUSE Client
在这个demo中我们采用在多个独立的节点部署FUSE Client的方式, 首先拷贝文件, 并创建mount点
cp ~/3fs/build/bin/hf3fs_fuse_main /opt/3fs/bin
cp ~/3fs/configs/{hf3fs_fuse_main_launcher.toml,hf3fs_fuse_main.toml,hf3fs_fuse_main_app.toml} /opt/3fs/etcmkdir -p /3fs/stage修改/opt/3fs/etc/hf3fs_fuse_main_launcher.toml配置如下:
cluster_id stage
mountpoint /3fs/stage
token_file /opt/3fs/etc/token.txt[mgmtd_client]
mgmtd_server_addresses [RDMA://10.99.0.1:8000]修改/opt/3fs/etc/hf3fs_fuse_main.toml配置如下
[mgmtd]
mgmtd_server_addresses [RDMA://10.99.0.1:8000][common.monitor.reporters.monitor_collector]
remote_ip 10.99.0.1:10000更新Fuse client配置到mgmtd service
/opt/3fs/bin/admin_cli -cfg /opt/3fs/etc/admin_cli.toml --config.mgmtd_client.mgmtd_server_addresses [RDMA://10.99.0.1:8000] set-config --type FUSE --file /opt/3fs/etc/hf3fs_fuse_main.toml开启fuse client
cp ~/3fs/deploy/systemd/hf3fs_fuse_main.service /usr/lib/systemd/system
systemctl start hf3fs_fuse_mainroot3fs-client:/opt/3fs# systemctl status hf3fs_fuse_main
● hf3fs_fuse_main.service - fuse_main ServerLoaded: loaded (/lib/systemd/system/hf3fs_fuse_main.service; disabled; vendor preset: enabled)Active: active (running) since Fri 2025-02-28 23:38:18 CST; 5s agoMain PID: 19841 (hf3fs_fuse_main)Tasks: 49 (limit: 629145)Memory: 318.9MCPU: 250msCGroup: /system.slice/hf3fs_fuse_main.service├─19841 /opt/3fs/bin/hf3fs_fuse_main --launcher_cfg /opt/3fs/etc/hf3fs_fuse_main_launcher.toml└─19903 fusermount3 --auto-unmount -- /3fs/stage
检查是否mount
root3fs-client:/opt/3fs# mount | grep /3fs/stage
hf3fs.stage on /3fs/stage type fuse.hf3fs (rw,nosuid,nodev,relatime,user_id0,group_id0,default_permissions,allow_other,max_read1048576)root3fs-meta:/opt/3fs# df -kh
Filesystem Size Used Avail Use% Mounted on
tmpfs 100G 1.9M 100G 1% /run
/dev/nvme0n1p3 394G 28G 350G 8% /
tmpfs 496G 16K 496G 1% /dev/shm
tmpfs 5.0M 0 5.0M 0% /run/lock
/dev/nvme0n1p2 197M 6.1M 191M 4% /boot/efi
tmpfs 100G 4.0K 100G 1% /run/user/0
hf3fs.stage 140T 999G 139T 1% /3fs/stage4. 性能测试
我们在5个fuse client上同时进行并发读取测试
fio -numjobs128 -fallocatenone -iodepth2 -ioenginelibaio -direct1 -rwread -bs4M --group_reporting -size100M -time_based -runtime3000 -name2depth_128file_4M_direct_read_bw -directory/3fs/stagedepth_128file_4M_direct_read_bw: (groupid0, jobs128): err 0: pid11785: Sat Mar 1 13:08:54 2025read: IOPS2669, BW10.4GiB/s (11.2GB/s)(6931GiB/664647msec) ##带宽为11.2GiB/s已经打满实例规格速度slat (usec): min36, max459933, avg47946.24, stdev11724.76clat (usec): min1303, max459937, avg47945.69, stdev11728.42lat (usec): min1891, max518800, avg95892.19, stdev16777.22clat percentiles (msec):| 1.00th[ 24], 5.00th[ 27], 10.00th[ 36], 20.00th[ 37],| 30.00th[ 47], 40.00th[ 48], 50.00th[ 49], 60.00th[ 50],| 70.00th[ 51], 80.00th[ 59], 90.00th[ 62], 95.00th[ 66],| 99.00th[ 79], 99.50th[ 86], 99.90th[ 97], 99.95th[ 102],| 99.99th[ 184]bw ( MiB/s): min 6192, max13702, per100.00%, avg10681.29, stdev 7.26, samples170112iops : min 1548, max 3422, avg2669.52, stdev 1.81, samples170112lat (msec) : 20.01%, 40.01%, 100.01%, 200.41%, 5069.00%lat (msec) : 10030.51%, 2500.05%, 5000.01%cpu : usr0.00%, sys0.18%, ctx6960833, majf0, minf363857IO depths : 10.1%, 2100.0%, 40.0%, 80.0%, 160.0%, 320.0%, 640.0%submit : 00.0%, 4100.0%, 80.0%, 160.0%, 320.0%, 640.0%, 640.0%complete : 00.0%, 4100.0%, 80.0%, 160.0%, 320.0%, 640.0%, 640.0%issued rwts: total1774252,0,0,0 short0,0,0,0 dropped0,0,0,0latency : target0, window0, percentile100.00%, depth2通过ECS管理控制台也可以看到已经打满带宽. 3FS还使用了clickhouse对运行数据进行统计分析, 可以登陆meta节点的查询
clickhouse-client --password eRDMA123!!
3fs-meta :) use 3fs3fs-meta :) select * from distributions where metricNamestorage_client.request_bw AND host3fs-fuse limit 10SELECT *
FROM distributions
WHERE (metricName storage_client.request_bw) AND (host 3fs-fuse)
LIMIT 10Query id: bae763a9-0c4c-413f-9103-a1c7fadaab6c┌───────────TIMESTAMP─┬─metricName────────────────┬─host─────┬─tag─┬─count─┬──────────────mean─┬────────────────min─┬────────────────max─┬───────────────p50─┬───────────────p90─┬────────────────p95─┬────────────────p99─┬─mount_name─┬─instance──┬─io─┬─uid─┬─method─┬─pod──────┬─thread─┬─statusCode─┐1. │ 2025-03-01 11:06:46 │ storage_client.request_bw │ 3fs-fuse │ │ 8591 │ 613373090.8395855 │ 216067587.0595508 │ 1675041533.546326 │ 594001794.1159781 │ 917471066.1935523 │ 1118230720.8216615 │ 1353937621.2941322 │ │ batchRead │ │ │ │ 3fs-fuse │ │ │2. │ 2025-03-01 11:06:47 │ storage_client.request_bw │ 3fs-fuse │ │ 10592 │ 631580288.6128079 │ 169261662.63115415 │ 1558062407.1322436 │ 625856561.6014094 │ 929167946.9438102 │ 1101134592.7515178 │ 1319381573.1463842 │ │ batchRead │ │ │ │ 3fs-fuse │ │ │3. │ 2025-03-01 11:06:48 │ storage_client.request_bw │ 3fs-fuse │ │ 10627 │ 624043291.1181132 │ 171476042.5183974 │ 1625699224.8062015 │ 620531561.6303563 │ 914151142.2446904 │ 1070662224.208625 │ 1305868148.0882857 │ │ batchRead │ │ │ │ 3fs-fuse │ │ │4. │ 2025-03-01 11:06:49 │ storage_client.request_bw │ 3fs-fuse │ │ 10660 │ 623494914.128004 │ 186214881.9037471 │ 1628223602.484472 │ 616674787.4039807 │ 914974468.4772394 │ 1089693011.0233178 │ 1281664891.6561015 │ │ batchRead │ │ │ │ 3fs-fuse │ │ │5. │ 2025-03-01 11:06:50 │ storage_client.request_bw │ 3fs-fuse │ │ 10627 │ 624230580.0179524 │ 221218565.4008439 │ 1620673879.4435859 │ 618548223.0963331 │ 918910082.9235835 │ 1088660510.2403255 │ 1292470925.9244142 │ │ batchRead │ │ │ │ 3fs-fuse │ │ │6. │ 2025-03-01 11:06:51 │ storage_client.request_bw │ 3fs-fuse │ │ 10606 │ 632341527.2096547 │ 185621525.9337936 │ 1605782542.1133232 │ 626439939.7497075 │ 928116483.5742279 │ 1114311679.0423079 │ 1318773172.0729723 │ │ batchRead │ │ │ │ 3fs-fuse │ │ │7. │ 2025-03-01 11:06:52 │ storage_client.request_bw │ 3fs-fuse │ │ 10591 │ 622737514.2896469 │ 176706437.47893494 │ 1596006088.2800608 │ 617361406.2508819 │ 910876455.9893316 │ 1076940139.2025425 │ 1297442965.0532806 │ │ batchRead │ │ │ │ 3fs-fuse │ │ │8. │ 2025-03-01 11:06:53 │ storage_client.request_bw │ 3fs-fuse │ │ 10635 │ 625666059.2437743 │ 188558892.28556016 │ 1600879389.312977 │ 619899922.4441694 │ 919915373.7100124 │ 1081999833.3945801 │ 1277120376.2726321 │ │ batchRead │ │ │ │ 3fs-fuse │ │ │9. │ 2025-03-01 11:06:54 │ storage_client.request_bw │ 3fs-fuse │ │ 10626 │ 622894588.6999174 │ 193001288.42260262 │ 1635843993.7597504 │ 618193735.7544193 │ 915741041.3424696 │ 1083785446.6478138 │ 1283924101.3262858 │ │ batchRead │ │ │ │ 3fs-fuse │ │ │
10. │ 2025-03-01 11:06:55 │ storage_client.request_bw │ 3fs-fuse │ │ 10622 │ 618279646.1477239 │ 200684401.9138756 │ 1598439024.390244 │ 610117422.3305147 │ 911149154.5515901 │ 1089556042.1725667 │ 1282777764.6938653 │ │ batchRead │ │ │ │ 3fs-fuse │ │ │└─────────────────────┴───────────────────────────┴──────────┴─────┴───────┴───────────────────┴────────────────────┴────────────────────┴───────────────────┴───────────────────┴────────────────────┴────────────────────┴────────────┴───────────┴────┴─────┴────────┴──────────┴────────┴────────────┘其它Metric可以通过如下命令查询3fs-meta :) select distinct metricName from distributionsSELECT DISTINCT metricName
FROM distributionsQuery id: c035f3af-9c97-4203-9d1b-ffee6eeeea44┌─metricName──────────────────────────────────────┐1. │ MgmtdClient.op.succ_latency │2. │ common_net_batch_read_size │3. │ common_net_batch_write_size │4. │ storage.check_disk.succ_latency │5. │ fdb_latency_commit │6. │ fdb_latency_get │7. │ fdb_latency_get_range │8. │ fdb_latency_snapshot_get_range │9. │ MgmtdService.WriterLatency │10. │ MgmtdService.bg.succ_latency │11. │ MgmtdService.op.succ_latency │12. │ storage.default.queue_latency │13. │ storage_client.concurrent_user_calls │14. │ storage_client.inflight_requests │15. │ storage_client.inflight_time │16. │ storage_client.network_latency │17. │ storage_client.num_pending_ops │18. │ storage_client.overall_latency │19. │ storage_client.request_latency │20. │ storage_client.server_latency │21. │ storage.io_submit.size │22. │ storage.io_submit.succ_latency │23. │ storage.read.queue_latency │24. │ storage.read_prepare_buffer.succ_latency │
这应该是全网首个复现3FS集群的测试, eRDMA提供的标准RDMA RC接口和全地域全可用区的弹性能力是我们能够快速复现的根本原因, 并且在云上可以根据用户需求构建更大规模的集群, 在ECS 9代服务器支持CIPU 2.0 400Gbps的处理能力及云上更大规模的资源供给能力下, 可以媲美DeepSeek线下部署的集群, 进一步优化推理的成本.
后续我们将针对3FS进行更多的测试和分析, 敬请期待~ 也希望这篇文章和阿里云eRDMA技术能够帮助您快速构建测试环境.
参考资料
[1]
ffrecord: https://github.com/HFAiLab/ffrecord
[2]
smallpond: https://github.com/deepseek-ai/smallpond
