为算材料、算化学、算物理的人而做Built for computational materials, chemistry and physics
不用再开终端敲 squeue、nvidia-smi、tail loss.out。Stop typing squeue, nvidia-smi and tail loss.out all day.
每台服务器一张卡片One card per server
鼠标移到刘海上就展开。集群和工作站分开显示，每张卡片上有自己的运行数、排队数、CPU / 内存 / GPU。没有任务的机器自动合并成一行，不占地方。 Hover the notch to expand. Clusters and workstations each get a card with their own running / pending counts, CPU, memory and GPUs. Idle servers fold into a single line.
- 读取 ~/.ssh/config，跳板机、密钥都照常生效Reads ~/.ssh/config — ProxyJump and keys just work
- 动态口令集群：输入一次，连接一直保持OTP clusters: log in once, stay connected
- 每台机器可以单独暂停、静音、设轮询间隔Pause, mute or set the interval per server
Slurm：数组、排队原因、分区空闲Slurm: arrays, pending reasons, free nodes
几十个元素的数组作业合并成一行，绿 / 红 / 蓝分段显示完成、失败、运行中。排队原因翻译成人话；需要你处理的（比如依赖永远无法满足）会标红。还能按需查看各分区还有多少空闲 CPU 和 GPU。 Job arrays collapse into one row with done / failed / running segments. Pending reasons are explained in plain words, and the ones that need you are flagged. Check free CPUs and GPUs per partition on demand.
完成和失败，刘海会告诉你The notch tells you when it's done
任务结束时刘海弹出提示，成功和失败是两种不同的提示音。没有开启作业记录的集群，也会按计算软件的输出判断是正常结束还是提前退出，不靠猜。 When a job ends, the notch pops up — with different sounds for success and failure. Even on clusters without job accounting, ServerIsland checks the software's own output to tell a clean finish from an early exit.
点开就是详情和日志末尾Details and log tail, one click away
作业信息、工作目录、实时刷新的日志末尾，一键复制路径或 tail -f 命令。取消任务默认关闭；打开后也要长按、二次确认，执行前还会在服务器上再核对一遍。 Job info, working directory and a live log tail; copy the path or a tail -f command in one click. Cancelling is off by default — and even when on, it takes a long press, a confirmation and a server-side re-check.