Add GPU scheduler plugin

wgs.py:
Weighted GPU scheduler, find a GPU_i that satisfies the minimum
(W_a * G_ui + W_b * G_mi)
W_a : default is 0.5
W_b : default is 0.5
G_ui: GPU utilization
G_mi: GPU memory usage

May assign to the same GPU when GPUs are always busy.

gfs.py:
Gfair scheduler

for (i = 0 to GPU_num-1)
	if (G[i].u < 1/W_a)
		G[i].u = 1/W_a

	G[i].cap = W_a*G[i].u + W_b*G[i].m

v_gpu[] = load gpu capability table

for (i = 0 to GPU_num-1)
	if v_gpu[i] = 0
		v_gpu[i] = 1.0

	G[i].cap = G[i].cap * v_gpu[i]

Find a G[k].cap that satisfies the minimum of {G[].cap}

v_gpu[k] = G[k].cap

atomic write v_gpu[] to the gpu capability table

Author: coldfunction

gfs.py

@@ -0,0 +1,99 @@
+from atomicwrites import atomic_write
+from subprocess import check_output, CalledProcessError
+
+class Utilization:
+    def __init__(self, gpu, memory):
+        self.gpu    = float(gpu)
+        self.memory = float(memory)
+        if self.gpu < 2:
+            self.gpu = 2
+    def set_gpu(self, gpu):
+        self.gpu    = gpu
+    def set_mem(self, memory):
+        self.memory = memory
+    def set_id(self, id):
+        self.id     = id
+    def set_cap(self, wa, wb):
+        self.cap    = wa*self.gpu + wb*self.memory
+
+output_gpu = check_output('nvidia-smi --query-gpu=utilization.gpu --format=csv', shell=True)
+output_gpu_split = output_gpu.split('\n')
+device_num = len(output_gpu_split) - 2
+
+d_gpu = []
+for i in range(device_num):
+    d_gpu.append(filter(str.isdigit, output_gpu_split[i+1]))
+    #print d_gpu[i]
+
+output_memory = check_output('nvidia-smi --query-gpu=memory.used --format=csv', shell=True)
+output_memory_split = output_memory.split('\n')
+
+d_memory = []
+for i in range(device_num):
+    d_memory.append(filter(str.isdigit, output_memory_split[i+1]))
+    #print d_memory[i]
+
+
+output_memory = check_output('nvidia-smi --query-gpu=memory.total --format=csv', shell=True)
+output_memory_split = output_memory.split('\n')
+
+for i in range(device_num):
+    d_memory[i] = float(d_memory[i]) / float(filter(str.isdigit, output_memory_split[i+1]))
+    #print d_memory[i]
+
+Wa = 0.5
+Wb = 0.5
+device_obj=[]
+for i in range(device_num):
+    device_obj.append(Utilization(d_gpu[i], d_memory[i]))
+    device_obj[i].set_cap(Wa, Wb)
+    device_obj[i].set_id(i)
+    #print device_obj[i].gpu
+    #print device_obj[i].memory
+    #print device_obj[i].cap
+    #print device_obj[i].id
+
+v_gpu = []
+f_gpu = []
+file = open('/home/coldfunction/qCUDA_0.1/qCUDA/.gpu_info', 'r')
+for i in range(device_num):
+    line = file.readline()
+    num = float(line)
+    
+    v_gpu.append(num)
+    f_gpu.append(num)
+#    if num == 0:
+#        print(num)
+file.close()
+
+
+for i in range(device_num):
+    if v_gpu[i] == 0:
+        v_gpu[i] = 1.0 
+   
+    device_obj[i].cap = (v_gpu[i] * device_obj[i].cap)
+    #print (device_obj[i].cap)
+
+    #print(v_gpu[i])
+    
+
+#print
+device_obj.sort(key=lambda i: i.cap) 
+
+id = device_obj[0].id
+
+f_gpu[id] = device_obj[0].cap
+
+with atomic_write('/home/coldfunction/qCUDA_0.1/qCUDA/.gpu_info', overwrite=True) as f:
+    for i in range(device_num):
+        f.write(str(f_gpu[i]))
+        f.write('\n')
+
+print id
+
+
+
+
+#f = open(".select_g", 'w')
+#s = str(device_obj[0].id)
+#f.write(s)

qcu-device/hw/misc/virtio-qcuda.c

@@ -5,7 +5,6 @@
 #include "hw/virtio/virtio-bus.h"
 #include "hw/virtio/virtio-qcuda.h"
 #include <sys/mman.h>
-//#include <stdio.h>
 
 #ifdef CONFIG_CUDA
 #include <cuda.h>
@@ -224,7 +223,7 @@ static void qcu_cudaRegisterFatBinary(VirtioQCArg *arg)
 	FILE *fp;
 	char buffer[20];
 	int id = 0;
-	fp=popen("python /home/cocotion/qcuda/select_gpu.py", "r");
+	fp=popen("python /home/coldfunction/qCUDA_0.1/qCUDA/gfs.py", "r");
 	id = (fgets(buffer, sizeof(buffer), fp) != NULL)?atoi(buffer):0;
 
 	pclose(fp);

wgs.py

@@ -0,0 +1,61 @@
+from subprocess import check_output, CalledProcessError
+
+class Utilization:
+    def __init__(self, gpu, memory):
+        self.gpu    = float(gpu)
+        self.memory = float(memory)
+    def set_gpu(self, gpu):
+        self.gpu    = gpu
+    def set_mem(self, memory):
+        self.memory = memory
+    def set_id(self, id):
+        self.id     = id
+    def set_cap(self, wa, wb):
+        self.cap    = wa*self.gpu + wb*self.memory
+
+output_gpu = check_output('nvidia-smi --query-gpu=utilization.gpu --format=csv', shell=True)
+output_gpu_split = output_gpu.split('\n')
+device_num = len(output_gpu_split) - 2
+
+d_gpu = []
+for i in range(device_num):
+    d_gpu.append(filter(str.isdigit, output_gpu_split[i+1]))
+    #print d_gpu[i]
+
+output_memory = check_output('nvidia-smi --query-gpu=memory.used --format=csv', shell=True)
+output_memory_split = output_memory.split('\n')
+
+d_memory = []
+for i in range(device_num):
+    d_memory.append(filter(str.isdigit, output_memory_split[i+1]))
+    #print d_memory[i]
+
+
+output_memory = check_output('nvidia-smi --query-gpu=memory.total --format=csv', shell=True)
+output_memory_split = output_memory.split('\n')
+
+for i in range(device_num):
+    d_memory[i] = float(d_memory[i]) / float(filter(str.isdigit, output_memory_split[i+1]))
+    #print d_memory[i]
+
+Wa = 0.5
+Wb = 0.5
+device_obj=[]
+for i in range(device_num):
+    device_obj.append(Utilization(d_gpu[i], d_memory[i]))
+    device_obj[i].set_cap(Wa, Wb)
+    device_obj[i].set_id(i)
+    #print device_obj[i].gpu
+    #print device_obj[i].memory
+    #print device_obj[i].cap
+    #print device_obj[i].id
+
+#print
+device_obj.sort(key=lambda i: i.cap) 
+#print device_obj[0].cap
+print device_obj[0].id
+
+
+#f = open(".select_g", 'w')
+#s = str(device_obj[0].id)
+#f.write(s)