AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

2.4.2-1

Browse Source 
Add tree algorithms for allreduce to improve performance at scale.
Add ncclCommAbort() and ncclCommGetAsyncError() to properly handle
network errors and be permit recover.
Detect initial CPU affinity and no longer escape it.
This commit is contained in:
Sylvain Jeaugey
2018-12-13 15:56:12 -08:00
parent 4861e197fd
commit 1450d42675
66 changed files with 3746 additions and 3251 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/include/transport.h
+49 -38
View File
@@ -1,5 +1,5 @@
/*************************************************************************
* Copyright (c) 2016-2018, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2016-2019, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
@@ -9,6 +9,7 @@
#include "nccl.h"
#include <stdint.h>
#include "nvmlwrap.h"
#define NTRANSPORTS 3
@@ -19,11 +20,13 @@ struct ncclRing;
struct ncclConnector;
struct ncclComm;
#define RANK_INFO_SIZE 64
typedef char ncclTinfo_t[RANK_INFO_SIZE];
struct ncclInfo {
ncclTinfo_t tinfo[NTRANSPORTS];
struct ncclPeerInfo {
int rank;
int cudaDev;
int nvmlDev;
uint64_t hostHash;
uint64_t pidHash;
char busId[NVML_DEVICE_PCI_BUS_ID_BUFFER_SIZE];
};
// Used to hold the transport connection values
@@ -34,18 +37,47 @@ struct ncclConnect {
char data[CONNECT_SIZE];
};
enum ncclProxyOpState { ncclProxyOpNone, ncclProxyOpReady, ncclProxyOpProgress, ncclProxyOpDone };
struct ncclProxyArgs;
typedef ncclResult_t (*proxyProgressFunc_t)(struct ncclProxyArgs*);
struct ncclProxyArgs {
struct ncclRing* ring;
int substeps;
proxyProgressFunc_t progress;
struct ncclChannel* channel;
struct ncclConnector* connector;
int sliceSteps;
int chunkSteps;
int nsteps;
uint64_t opCount;
int llMode;
bool needProxy;
int active; // add component before this line -- it is left out during initialization
int state; // add component before this line -- it is left out during initialization
// Internal state
uint64_t head;
uint64_t tail;
uint64_t end;
void* requests[NCCL_STEPS];
int idle;
// Element linking
pthread_mutex_t mutex;
struct ncclProxyArgs* next;
struct ncclProxyArgs* nextPeer;
};
struct ncclProxyPool;
struct ncclProxyState {
pthread_cond_t cond;
pthread_mutex_t mutex;
bool stop;
struct ncclProxyArgs* ops;
struct ncclProxyArgs* pool;
struct ncclProxyPool* pools;
};
struct ncclTransportComm {
ncclResult_t (*setup)(ncclTinfo_t*, ncclTinfo_t*, struct ncclConnect*, struct ncclRing*);
ncclResult_t (*setup)(struct ncclPeerInfo*, struct ncclPeerInfo*, struct ncclConnect*, struct ncclConnector*, int buffSize, int channelId);
ncclResult_t (*connect)(struct ncclConnect*, struct ncclConnector*);
ncclResult_t (*free)(void*);
ncclResult_t (*proxy)(struct ncclProxyArgs*);
@@ -53,8 +85,7 @@ struct ncclTransportComm {
struct ncclTransport {
const char name[4];
ncclResult_t (*fillInfo)(ncclTinfo_t*, int);
ncclResult_t (*canConnect)(ncclTvalue_t*, ncclTinfo_t*, ncclTinfo_t*);
ncclResult_t (*canConnect)(ncclTvalue_t*, struct ncclPeerInfo*, struct ncclPeerInfo*);
ncclResult_t (*getRings)(int, int*, int*, ncclTvalue_t*, int*, int*, int*, int, int*);
struct ncclTransportComm send;
struct ncclTransportComm recv;
@@ -64,37 +95,17 @@ struct ncclTransport {
typedef ncclResult_t (*threadFunc_t)(struct ncclProxyArgs*);
#define TRANSPORT_PROXY_FIFO_SIZE NCCL_MAX_OPS
struct transportProxyInfo {
struct ncclComm* comm;
pthread_t thread;
threadFunc_t func;
volatile int proxyReady;
struct ncclProxyArgs argsFifo[TRANSPORT_PROXY_FIFO_SIZE];
volatile uint64_t argsFifoHead;
volatile uint64_t argsFifoTail;
pthread_cond_t cond;
pthread_mutex_t mutex;
};
ncclResult_t transportCreateProxy(int type, struct ncclRing* ring, struct ncclComm* comm);
ncclResult_t transportDestroyProxy(struct ncclConnector* connector);
enum proxyMode {
proxyRing = 0,
proxyFrom = 1,
proxyTo = 2
};
static int proxyPatternRing = proxyRing;
static inline int proxyPatternFrom(int root) { return 1+root; }
static inline int proxyPatternTo(int root) { return -1-root; }
static inline enum proxyMode proxyPatternMode(int pattern) { return (pattern == 0) ? proxyRing : ((pattern > 0) ? proxyFrom : proxyTo); }
static inline int proxyPatternRoot(int pattern) { return (pattern > 0) ? pattern-1 : -pattern-1; }
ncclResult_t transportSaveProxies(int substeps, int subchunks, int nstepsPerRound, int nblocksPerRound, size_t size, int pattern, struct ncclComm* comm);
ncclResult_t transportStartProxies(struct ncclComm* comm);
ncclResult_t transportAllocateProxyArgs(struct ncclComm* comm, struct ncclProxyArgs** argsptr);
ncclResult_t transportSaveProxies(struct ncclProxyArgs* args, int pattern, int root, int nranks);
ncclResult_t transportStartProxy(struct ncclComm* comm);
ncclResult_t transportCreateProxy(struct ncclComm* comm);
ncclResult_t transportDestroyProxy(struct ncclComm* comm);
#include <unistd.h>