update some links

docs/api.md

@@ -27,7 +27,7 @@ often server *i* handles the keys (feature indices) within the i-th
 
 , which is often used to monitor and control the
   progress of the machine learning application. It also can be used to deal with node
-  failures. See an example in [asynchronous SGD](https://github.com/dmlc/wormhole/blob/master/learn/solver/async_sgd.h#L27).
+  failures. See an example in [asynchronous SGD](https://github.com/dmlc/wormhole/blob/master/learn/linear/async_sgd.h#L27).
 
 ```eval_rst
 .. automodule:: ps::KVWorker

docs/env.md

@@ -14,4 +14,4 @@ additional variables:
   automatically
 - `DMLC_LOCAL` : runs in local machines, no network is needed
 - `DMLC_PS_WATER_MARK`  : limit on the maximum number of outstanding messages
-- `DMLC_PS_VAN_TYPE` : the type of the Van for transport, can be `ibverbs` for RDMA, `zmq` for TCP, `p3` for TCP with [priority based parameter propagation](https://anandj.in/wp-content/uploads/sysml.pdf).
+- `DMLC_PS_VAN_TYPE` : the type of the Van for transport, can be `ibverbs` for RDMA, `zmq` for TCP, `p3` for TCP with [priority based parameter propagation](https://proceedings.mlsys.org/paper/2019/hash/d09bf41544a3365a46c9077ebb5e35c3-Abstract.html).

docs/overview.md

@@ -177,6 +177,6 @@ name some of them
   convergence rate with large minibatch size *b*
 - [Sra, AISTATS'16](http://arxiv.org/abs/1508.05003) asynchronous SGD adaptive
   to the actually delay rather than the worst maximal delay
-- [Li, WSDM'16](http://www.cs.cmu.edu/~yuxiangw/docs/fm.pdf) practical
+- [Li, WSDM'16](https://sites.cs.ucsb.edu/~yuxiangw/docs/fm.pdf) practical
   considerations for asynchronous SGD with the parameter server
 - [Chen, LearningSys'16]() synchronized SGD for deep learning.