Summary of Env Variables#

Summary of Basic Environment Variables Using by GeoMX.#

Environment Variable

Options

Node Used On

Description

DMLC_ROLE

scheduler, server, worker

local scheduler, local server, master worker, worker

The role of the node within the party.

DMLC_PS_ROOT_URI

IPv4 address

global server, master worker, local scheduler, local server, worker

IPv4 address of the local scheduler node.

DMLC_PS_ROOT_PORT

Integer

same as above

Port number of the local scheduler node.

DMLC_NUM_SERVER

Integer

same as above

Number of local servers in the participating party, or number of global servers in the central party.

DMLC_NUM_WORKER

Integer

same as above

Number of workers in the current party, including the master worker.

DMLC_ROLE_GLOBAL

global_scheduler, global_server

global scheduler, global server

The role of the node across different parties.

DMLC_PS_GLOBAL_ROOT_URI

IPv4 address

global scheduler, global server, local server

IPv4 address of the global scheduler node.

DMLC_PS_GLOBAL_ROOT_PORT

Integer

same as above

Port number of the global scheduler node.

DMLC_NUM_GLOBAL_SERVER

Integer

same as above

Number of global servers in the central party.

DMLC_NUM_GLOBAL_WORKER

Integer

same as above

Number of local servers worldwide.

DMLC_ROLE_MASTER_WORKER

0, 1

master worker

Specify if the current node is the master worker.

DMLC_ENABLE_CENTRAL_WORKER

0, 1

global server

Specify if the central party joins in model training.

DMLC_NUM_ALL_WORKER

Integer

global server, master worker, worker

Total number of workers actually participating in model training.

DMLC_INTERFACE

String

all

Name of the network interface used by the node.

PS_VERBOSE

0, 1, 2

all

Verbosity level of the system logs.

Summary of Environment Variables for Each Optimization Technology.#

Technology

Environment Variable

Description

HFA

MXNET_KVSTORE_USE_HFA

Enable or disable HFA synchronization.

MXNET_KVSTORE_HFA_K1

Number of loops before a local synchronization.

MXNET_KVSTORE_HFA_K2

Number of loops before a global synchronization.

Bi-Sparse, MPQ

MXNET_KVSTORE_SIZE_LOWER_BOUND

Size lower bound for classifying large and tiny tensors.

DGT

ENABLE_DGT

Enable or disable DGT protocol, set to 2 for enable.

DMLC_UDP_CHANNEL_NUM

Number of channels used for transmission.

DMLC_K

Compression ratio.

ADAPTIVE_K_FLAG

Enable or disable adaptive K value adjustment.

DGT_CONTRIBUTION_ALPHA

The weighted average coefficient used to update gradient contributions.

DGT_INFO

Enable or disable the debug information of DGT.

DGT_BLOCK_SIZE

The block size for gradient classification, default is 4096.

DMLC_K_MIN

The lower limit of DMLC_K when ADAPTIVE_K_FLAG is enabled,default is 0.2.

TSEngine

ENABLE_INTER_TS

Enable or disable TSEngine within the data center.

ENABLE_INTRA_TS

Enable or disable TSEngine between data centers.

MAX_GREED_RATE_TS

Probability set for random exploration.

P3

ENABLE_P3

Enable or disable P3 scheduler.