Each trainer runs the forward and backward passes using their local
data:
1. In the forward pass, when a trainer runs the forward algorithm of a
lookup operator, it retrieves W(x) from the storage service.
1. The trainer computes W'(x) in the backward pass using W(x).
During the global update process:
1. Each trainer uploads its W'(x) to parameter servers.
1. The parameter server runs the optimization algorithm, e.g., the
Adam optimization algorithm, which requires that
1. The parameter server retrieves W(x) from memcached, and
1. The parameter server pushes $\Delta W(x)=f(W(x), lambda \sum_j
W'(x))$ to memcached, where $f$ denotes the optimization
algorithm.
### Storage Service Does Optimize
This design is very similar to the above one, except that the
optimization algorithm $f$ runs on the storage service.
- Pro: parameter servers do not retrieve W(x) from the storage
service, thus saves half network communication.
- Con: the storage service needs to be able to run the optimization
algorithm.
## Conclusion
Let us do the "storage service does not optimize" solution first, as a
baseline at least, because it is easier to use a well-optimized
distributed storage service like memcached. We can do the "storage
service does optimize" solution later or at the same time, which, if
implemented carefully, should have better performance than the former.
## Distributed Lookup Table
### Problem 1: The lookup table may be very large.
In the condition like the search engine and recommendation system, the number of feature Id may be very large, say 100,000,000,000, then for a float value lookup table of size 8, the total size of the table is:
```
100,000,000,000 * 8 * 4(Bytes) = 2980.23 GB
```
### Solution: Distributed storage
1. Paddle use [SelectedRows](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/selected_rows.md) as the storage format for the lookup table, the lookup table parameter will be split to multi-machine according to the hash of the feature ID, and data will also be split and send to the same machine to prefetch the parameter.
1. For common parameters, the trainer will get the whole parameter for training, but for the big lookup table, the trainer can not store the whole parameter. Because the input data feature is very sparse, every time we only need a few parameters for training, so we use `prefetch_op` to only prefetch the parameter needed to trainer.
### Problem 2. The Id in the lookup table is not sure before training.
The feature Id is calculated by the hash function because the feature data source is so large, we can not get all the Id before training. So we can not initialize the table before training.
### Solution: Id auto growth
At the beginning of training, paddle only malloc the memory for the lookup table at parameter server side, the Id and it's value will not be initialized. During training, when a parameter server received an Id, if it is already in the lookup table, it will return the existing parameter, if the Id does not exist, paddle will add it into the lookup table and initialize the value for it.
### Problem 3: parameter load and save
For common parameters, paddle use trainer to save and load them. But for distributed lookup table, trainer cannot do this because it's large size.
### Solution: Parameter server side save and load
Paddle support parameter server side save and load for distribute lookup table. Each machine of parameter servers will only save and load part of the whole table.
## Architecture
The whole architecture of the distribute lookup table is as below:
### Training steps:
1. Read a batch of data, the data is feature ids.
1. The input ids will be split by `split_ids_op` with the same hash function of the lookup table.
1. The `prefetch_op` use the split result to prefetch parameters back from the lookup table.
1. Run forward-backward to get the gradient of the lookup table.
1. `split_ids_op` split the gradient and then use `send_op` to the parameter server.
1. parameter server update the table with the received gradient.
sneaxiy
7 years ago
