fix some problem

analysis.md

+对于热点数据的为k的通信比例是
+
+假设$y$作为跨分区访问的比例
+
+$N_{root}^k\cdot \eta_1\cdot m \cdot d_{mem} + (3d_{mem} + d_{e} + d_{n})\cdot \theta \cdot y \cdot N_{nei} $
+
+选择k带来的y可以先暂时估算为
+$((|E|-|E^k|)*y+|E^k|*\frac{m-1}{m})/|E|$
+
+化简一下远程访问的比例大概是$(1-|E^k|/|E|)*y + |E^k|/E*\frac{m-1}{m}$
+
+
+那么通信量可以表示为$|E^k|\cdot \eta_1\cdot m \cdot d_{mem} + d_{fetch}\cdot N_{nei} \cdot \theta \cdot ((1-|E^k|/|E|)+ |E^k|/E*\frac{m-1}{m})$
+
+化简后可以表示为
+$|E^k|(\eta_1\cdot m \cdot d_{mem} - d_{fetch} N_{nei}\cdot \theta/|E|) + d_{fetch} \cdot N_{nei} \cdot \theta$
+
+那么$\eta_1 < d_{fetch} N_{nei} \cdot \theta/|E|/({d_{mem} \cdot m}) $带来优势
+
+$\eta_1 < \frac{3}{m}N_{nei}/|E|$则能够减少通信的开销，实验验证当$\alpha$大于0.7的时候，$\eta_1$小于0.1
+
+
+
+从精度角度考虑，采用k共享之后，能够采样得到的本地特征的增加数量
+
+对于热点数据,访问外部分区的数量
+$|E_k| * \frac{(|E|-|E_k|)}{|E|}\frac{m-1}{m}$
+
+对于非热点数据，访问外部分区的数量
+$(|E|-|E_k|)*\frac{|E|-|E_k|}{E}y$

examples/test_all.sh

@@ -6,9 +6,9 @@ addr="192.168.1.107"
 partition_params=("ours")
 #"metis" "ldg" "random")
 #("ours" "metis" "ldg" "random")
-partitions="4"
+partitions="8"
 node_per="4"
-nnodes="1"
+nnodes="2"
 node_rank="0"
 probability_params=("0.1")
 sample_type_params=("boundery_recent_decay")
@@ -19,7 +19,7 @@ memory_type=("historical")
 #memory_type=("local" "all_update" "historical" "all_reduce")
 shared_memory_ssim=("0.3")
 #data_param=("WIKI" "REDDIT" "LASTFM" "WikiTalk")
-data_param=("WikiTalk")
+data_param=("LASTFM" "WikiTalk" "StackOverflow") 
 #"GDELT")
 #data_param=("WIKI" "REDDIT" "LASTFM" "DGraphFin" "WikiTalk" "StackOverflow")
 #data_param=("WIKI" "REDDIT" "LASTFM" "WikiTalk" "StackOverflow")
@@ -32,9 +32,9 @@ data_param=("WikiTalk")
 #seed=(( RANDOM % 1000000 + 1 ))
 mkdir -p all_"$seed"
 for data in "${data_param[@]}"; do
-    model="JODIE_large"
+    model="TGN_large"
     if [ "$data" = "WIKI" ] || [ "$data" = "REDDIT" ] || [ "$data" = "LASTFM" ]; then
-        model="JODIE"
+        model="TGN"
     fi
     #model="APAN"
     mkdir all_"$seed"/"$data"

starrygl/module/utils.py

@@ -137,7 +137,7 @@ class AdaParameter:
     self.alpha = max(min(self.alpha, self.max_alpha),self.min_alpha)
     print('gnn aggregate {} fetch {} memory sync {} memory update {}'.format(average_gnn_aggregate,average_fetch,average_memory_sync_time,average_memory_update_time))
     print('beta is {} alpha is {}\n'.format(self.beta,self.alpha))
-    self.reset_time()
+    #self.reset_time()
     #log（2-a1 ） = log(2-a2) * t1/t2 * (1 + wait_threshold)
     #2-a1 = 2-a2 ^(t1/t2 * (1 + wait_threshold))
     #a1 = 2 - 2-a2 ^(t1/t2 * (1 + wait_threshold))

starrygl/sample/data_loader.py

@@ -224,10 +224,10 @@ class DistributedDataLoader:
             next_data = self.input_dataset[torch.tensor([],device=self.device,dtype= torch.long)]
         else:
             next_data = self.input_dataset[self.batch_pos_l[self.submitted]:self.batch_pos_r[self.submitted] +1]
-        if self.mode=='train' and self.probability < 1:
+        if self.mode=='train' and self.ada_param.beta < 1:
             mask = ((DistIndex(self.graph.nids_mapper[next_data.edges[0,:].to('cpu')]).part == dist.get_rank())&(DistIndex(self.graph.nids_mapper[next_data.edges[1,:].to('cpu')]).part == dist.get_rank()))
-            if self.probability > 0:
-                mask[~mask] = (torch.rand((~mask).sum().item()) < self.probability)
+            if self.ada_param.beta > 0:
+                mask[~mask] = (torch.rand((~mask).sum().item()) < self.ada_param.beta)
             next_data = next_data[mask.to(next_data.device)]
         self.submitted = self.submitted + 1
         return next_data