278 lines
6.4 KiB
Markdown
278 lines
6.4 KiB
Markdown
如果升级到 **Fortran 2018**,方案会干净很多:可以把老的 Fortran95 计算核心保留,但用 Fortran 2018 的 `iso_c_binding`、`iso_fortran_env`、`event_type`/coarray 等现代能力做一层更稳的接口。不过要注意:**Fortran 2018 本身仍然没有标准的“操作系统共享内存 API”**,所以共享内存创建、映射、命名信号量这类事,仍建议由 C/C++ 或系统库封装。
|
||
|
||
我会推荐这个架构:
|
||
|
||
```text
|
||
Controller
|
||
进程启动 / 同步调度 / 状态机 / 异常恢复
|
||
|
|
||
control IPC: pipe/socket/semaphore/event
|
||
/ \
|
||
/ \
|
||
Solver A Solver B
|
||
Fortran 2018 Fortran 2018
|
||
计算 + IPC接口层 计算 + IPC接口层
|
||
|
||
shared memory A_to_B
|
||
shared memory B_to_A
|
||
```
|
||
|
||
核心原则不变:
|
||
|
||
```text
|
||
大数据:共享内存
|
||
小命令:控制通道
|
||
同步权:Controller
|
||
A/B:只按命令读写,不直接调度对方
|
||
```
|
||
|
||
**推荐模块划分**
|
||
|
||
```text
|
||
controller
|
||
- 创建共享内存
|
||
- 创建同步对象
|
||
- 启动 A/B
|
||
- 发送 step/read/write/stop 命令
|
||
- 等待 A/B done
|
||
- 处理超时、崩溃、重启
|
||
|
||
ipc_c_layer
|
||
- shm_create / shm_open / shm_map
|
||
- event_wait / event_signal
|
||
- mutex_lock / mutex_unlock
|
||
- pipe_send / pipe_recv
|
||
|
||
fortran_ipc_module
|
||
- Fortran 2018 包装层
|
||
- 使用 iso_c_binding 调 C 接口
|
||
- 把 C 指针转换成 Fortran 数组指针
|
||
|
||
solver_a / solver_b
|
||
- 原计算逻辑
|
||
- 调用 fortran_ipc_module 读写共享内存
|
||
```
|
||
|
||
**共享内存布局**
|
||
|
||
建议固定二进制布局,避免复杂对象:
|
||
|
||
```text
|
||
SharedBlock
|
||
Header
|
||
magic
|
||
version
|
||
role
|
||
step_id
|
||
nx, ny, nz
|
||
elem_kind
|
||
data_bytes
|
||
state
|
||
error_code
|
||
Payload
|
||
real(real64) data(nx, ny, nz)
|
||
```
|
||
|
||
Fortran 2018 侧可以通过 `c_f_pointer` 把共享内存 payload 映射成数组:
|
||
|
||
```fortran
|
||
use iso_c_binding
|
||
use iso_fortran_env, only: real64
|
||
|
||
type(c_ptr) :: raw_ptr
|
||
real(real64), pointer :: field(:,:,:)
|
||
|
||
raw_ptr = ipc_get_payload_ptr(handle)
|
||
call c_f_pointer(raw_ptr, field, [nx, ny, nz])
|
||
```
|
||
|
||
这样 Solver 看到的就是普通 Fortran 数组:
|
||
|
||
```fortran
|
||
call compute(field)
|
||
```
|
||
|
||
**双向数据区**
|
||
|
||
为了减少锁竞争,建议不要共用一块 buffer,而是两个单向通道:
|
||
|
||
```text
|
||
A_to_B_shm:
|
||
A 只写
|
||
B 只读
|
||
|
||
B_to_A_shm:
|
||
B 只写
|
||
A 只读
|
||
```
|
||
|
||
如果是实时流数据,再加双缓冲或三缓冲:
|
||
|
||
```text
|
||
A_to_B:
|
||
buffer[0]
|
||
buffer[1]
|
||
buffer[2]
|
||
```
|
||
|
||
Header 里放:
|
||
|
||
```text
|
||
write_index
|
||
read_index
|
||
latest_step
|
||
buffer_state[3]
|
||
```
|
||
|
||
**同步流程示例**
|
||
|
||
```text
|
||
Controller:
|
||
create shared memory
|
||
start A
|
||
start B
|
||
|
||
loop step = 1..N:
|
||
signal A: compute_and_write_A_to_B(step)
|
||
wait A_done
|
||
|
||
signal B: read_A_to_B_and_compute(step)
|
||
wait B_done
|
||
|
||
signal B: write_B_to_A(step)
|
||
wait B_done
|
||
|
||
signal A: read_B_to_A_and_update(step)
|
||
wait A_done
|
||
|
||
shutdown:
|
||
signal A/B stop
|
||
release resources
|
||
```
|
||
|
||
如果 A/B 可以并行计算,则改成:
|
||
|
||
```text
|
||
Controller:
|
||
signal A and B start_step
|
||
wait both done
|
||
signal both exchange_ready
|
||
wait both consumed
|
||
```
|
||
|
||
**Fortran 2018 带来的好处**
|
||
|
||
相比 Fortran95,主要提升在接口层:
|
||
|
||
| 能力 | Fortran95 | Fortran2018 |
|
||
|---|---|---|
|
||
| 调 C 接口 | 依赖编译器扩展 | 标准 `iso_c_binding` |
|
||
| 整数/浮点 kind | 不完全统一 | `iso_fortran_env` 更清楚 |
|
||
| C 指针转数组 | 麻烦且不标准 | `c_f_pointer` 标准支持 |
|
||
| 字符串/枚举/句柄 | 易混乱 | 可用 `c_ptr`, `c_int`, `c_size_t` |
|
||
| 并行同步 | 依赖外部库 | 可考虑 coarray/event,但跨进程仍有限 |
|
||
|
||
**是否使用 coarray?**
|
||
|
||
可以考虑,但我不建议作为第一方案。
|
||
|
||
Fortran 2018 的 coarray/event 适合 Fortran 程序之间的并行模型,但你的场景是:
|
||
|
||
- 两个独立主控程序
|
||
- 由第三个主控程序统一调度
|
||
- 大数据同机交换
|
||
- 需要工程上可控的进程生命周期
|
||
|
||
这种情况下,coarray 运行时可能反而把进程模型绑定到特定编译器和运行时,例如 Intel、GNU/OpenCoarrays、Cray 等。可移植性、调试和与外部 Controller 集成会更麻烦。
|
||
|
||
所以优先级我会排成:
|
||
|
||
```text
|
||
首选:Fortran 2018 + C IPC 封装 + 共享内存 + Controller
|
||
备选:MPI + Fortran 2018,如果你们本来就是 HPC/MPI 体系
|
||
谨慎:Fortran coarray,用于同一 Fortran 运行时生态内的并行程序
|
||
```
|
||
|
||
**最小接口设计**
|
||
|
||
C 层暴露给 Fortran 的接口可以保持很少:
|
||
|
||
```c
|
||
void* ipc_attach(const char* name, int role);
|
||
void* ipc_payload_ptr(void* handle, int buffer_id);
|
||
int ipc_wait_command(void* handle, int* command, int* step_id);
|
||
int ipc_signal_done(void* handle, int step_id, int status);
|
||
int ipc_detach(void* handle);
|
||
```
|
||
|
||
Fortran 2018 包一层:
|
||
|
||
```fortran
|
||
module ipc_mod
|
||
use iso_c_binding
|
||
implicit none
|
||
|
||
type :: ipc_handle
|
||
type(c_ptr) :: p = c_null_ptr
|
||
end type
|
||
|
||
contains
|
||
|
||
subroutine attach(handle, name, role)
|
||
end subroutine
|
||
|
||
subroutine map_real64_3d(handle, buffer_id, nx, ny, nz, field)
|
||
type(ipc_handle), intent(in) :: handle
|
||
integer, intent(in) :: buffer_id, nx, ny, nz
|
||
real(c_double), pointer, intent(out) :: field(:,:,:)
|
||
end subroutine
|
||
|
||
subroutine wait_command(handle, command, step_id)
|
||
end subroutine
|
||
|
||
subroutine signal_done(handle, step_id, status)
|
||
end subroutine
|
||
|
||
end module
|
||
```
|
||
|
||
Solver 主循环大概是:
|
||
|
||
```fortran
|
||
do
|
||
call wait_command(ipc, cmd, step_id)
|
||
|
||
select case (cmd)
|
||
case (CMD_COMPUTE_WRITE)
|
||
call compute(local_state, out_field)
|
||
call signal_done(ipc, step_id, STATUS_OK)
|
||
|
||
case (CMD_READ_UPDATE)
|
||
call update_from_peer(in_field)
|
||
call signal_done(ipc, step_id, STATUS_OK)
|
||
|
||
case (CMD_STOP)
|
||
exit
|
||
end select
|
||
end do
|
||
```
|
||
|
||
**最终推荐**
|
||
|
||
升级到 Fortran 2018 后,最合适的方案是:
|
||
|
||
```text
|
||
Controller:
|
||
独立主控程序,管理 A/B 生命周期和同步状态机
|
||
|
||
Solver A / Solver B:
|
||
Fortran 2018 独立程序
|
||
保留各自计算主循环
|
||
通过统一 ipc_mod 接收命令、读写共享内存、上报状态
|
||
|
||
IPC 底层:
|
||
C/C++ 封装 OS 共享内存 + semaphore/event/named pipe
|
||
Fortran 2018 用 iso_c_binding 标准调用
|
||
```
|
||
|
||
这比 Fortran95 的版本更可维护、可移植、可测试。关键是把系统相关的复杂度压进很薄的 C/C++ IPC 层,让 Fortran 代码只面对“数组 + 命令 + 状态”。 |