vitalif
/
mdtest
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity
mdtest/src/aiori-S3.c

1505 lines
48 KiB
C
Executable File
Raw Blame History

/* -*- mode: c; indent-tabs-mode: t; -*-
* vim:noexpandtab:
*
* Editing with tabs allows different users to pick their own indentation
* appearance without changing the file.
*/
/*
* Copyright (c) 2009, Los Alamos National Security, LLC All rights reserved.
* Copyright 2009. Los Alamos National Security, LLC. This software was produced
* under U.S. Government contract DE-AC52-06NA25396 for Los Alamos National
* Laboratory (LANL), which is operated by Los Alamos National Security, LLC for
* the U.S. Department of Energy. The U.S. Government has rights to use,
* reproduce, and distribute this software. NEITHER THE GOVERNMENT NOR LOS
* ALAMOS NATIONAL SECURITY, LLC MAKES ANY WARRANTY, EXPRESS OR IMPLIED, OR
* ASSUMES ANY LIABILITY FOR THE USE OF THIS SOFTWARE. If software is
* modified to produce derivative works, such modified software should be
* clearly marked, so as not to confuse it with the version available from
* LANL.
*
* Additionally, redistribution and use in source and binary forms, with or
* without modification, are permitted provided that the following conditions are
* met:
*
* • Redistributions of source code must retain the above copyright notice,
* this list of conditions and the following disclaimer.
*
* • Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* • Neither the name of Los Alamos National Security, LLC, Los Alamos National
* Laboratory, LANL, the U.S. Government, nor the names of its contributors may be
* used to endorse or promote products derived from this software without specific
* prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY LOS ALAMOS NATIONAL SECURITY, LLC AND CONTRIBUTORS
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
* THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL LOS ALAMOS NATIONAL SECURITY, LLC OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT
* OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
* IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY
* OF SUCH DAMAGE.
*/
/******************************************************************************
*
* Implementation of abstract IOR interface, for the Amazon S3 API.
* EMC/ViPR supports some useful extensions to S3, which we also implement
* here. There are 3 different mixes:
*
* (1) "Pure S3" uses S3 "Multi-Part Upload" to do N:1 writes. N:N writes
* fail, in the case where IOR "transfer-size" differs from
* "block-size', because this implies an "append", and append is not
* supported in S3. [TBD: The spec also says multi-part upload can't
* have any individual part greater than 5MB, or more then 10k total
* parts. Failing these conditions may produce obscure errors. Should
* we enforce? ]
*
* --> Select this option with the '-a S3' command-line arg to IOR
*
*
* (2) "EMC S3 Extensions" uses the EMC byte-range support for N:1
* writes, eliminating Multi-Part Upload. EMC expects this will
* perform better than MPU, and it avoids some problems that are
* imposed by the S3 MPU spec. [See comments at EMC_Xfer().]
*
* --> Select this option with the '-a EMC_S3' command-line arg to IOR
*
*
* NOTE: Putting EMC's S3-extensions in the same file with the S3 API
* allows us to share some code that would otherwise be duplicated
* (e.g. s3_connect(), etc). This should also help us avoid losing
* bug fixes that are discovered in one interface or the other. In
* some cases, S3 is incapable of supporting all the needs of IOR.
* (For example, see notes about "append", above S3_Xfer().
*
******************************************************************************/
#ifdef HAVE_CONFIG_H
# include "config.h"
#endif
#include <stdio.h>
#include <stdlib.h>
#include <string.h> /* strnstr() */
#include <errno.h>
#include <assert.h>
/*
#ifdef HAVE_LUSTRE_USER
#include <lustre/lustre_user.h>
#endif
*/
#include "ior.h"
#include "aiori.h"
#include "iordef.h"
#include <curl/curl.h>
#include <libxml/parser.h> // from libxml2
#include <libxml/tree.h>
#include "aws4c.h" // extended vers of "aws4c" lib for S3 via libcurl
#include "aws4c_extra.h" // utilities, e.g. for parsing XML in responses
/* buffer is used to generate URLs, err_msgs, etc */
#define BUFF_SIZE 1024
static char buff[BUFF_SIZE];
const int ETAG_SIZE = 32;
CURLcode rc;
/* Any objects we create or delete will be under this bucket */
const char* bucket_name = "ior";
/* TODO: The following stuff goes into options! */
/* REST/S3 variables */
// CURL* curl; /* for libcurl "easy" fns (now managed by aws4c) */
# define IOR_CURL_INIT 0x01 /* curl top-level inits were performed once? */
# define IOR_CURL_NOCONTINUE 0x02
# define IOR_CURL_S3_EMC_EXT 0x04 /* allow EMC extensions to S3? */
#ifdef USE_S3_AIORI
# include <curl/curl.h>
# include "aws4c.h"
#else
typedef void CURL; /* unused, but needs a type */
typedef void IOBuf; /* unused, but needs a type */
#endif
IOBuf* io_buf; /* aws4c places parsed header values here */
IOBuf* etags; /* accumulate ETags for N:1 parts */
///////////////////////////////////////////////
/**************************** P R O T O T Y P E S *****************************/
static void* S3_Create(char*, IOR_param_t*);
static void* S3_Open(char*, IOR_param_t*);
static IOR_offset_t S3_Xfer(int, void*, IOR_size_t*, IOR_offset_t, IOR_param_t*);
static void S3_Close(void*, IOR_param_t*);
static void* EMC_Create(char*, IOR_param_t*);
static void* EMC_Open(char*, IOR_param_t*);
static IOR_offset_t EMC_Xfer(int, void*, IOR_size_t*, IOR_offset_t, IOR_param_t*);
static void EMC_Close(void*, IOR_param_t*);
static void S3_Delete(char*, IOR_param_t*);
static void S3_Fsync(void*, IOR_param_t*);
static IOR_offset_t S3_GetFileSize(IOR_param_t*, MPI_Comm, char*);
static void S3_init(void * options);
static void S3_finalize(void * options);
static int S3_check_params(IOR_param_t *);
/************************** D E C L A R A T I O N S ***************************/
// "Pure S3"
// N:1 writes use multi-part upload
// N:N fails if "transfer-size" != "block-size" (because that requires "append")
ior_aiori_t s3_aiori = {
.name = "S3",
.name_legacy = NULL,
.create = S3_Create,
.open = S3_Open,
.xfer = S3_Xfer,
.close = S3_Close,
.delete = S3_Delete,
.get_version = aiori_get_version,
.fsync = S3_Fsync,
.get_file_size = S3_GetFileSize,
.initialize = S3_init,
.finalize = S3_finalize,
.check_params = S3_check_params
};
// "S3", plus EMC-extensions enabled
// N:1 writes use multi-part upload
// N:N succeeds (because EMC-extensions support "append")
ior_aiori_t s3_plus_aiori = {
.name = "S3_plus",
.create = S3_Create,
.open = S3_Open,
.xfer = S3_Xfer,
.close = S3_Close,
.delete = S3_Delete,
.set_version = S3_SetVersion,
.fsync = S3_Fsync,
.get_file_size = S3_GetFileSize,
.initialize = S3_init,
.finalize = S3_finalize
};
// Use EMC-extensions for N:1 write, as well
// N:1 writes use EMC byte-range
// N:N succeeds because EMC-extensions support "append"
ior_aiori_t s3_emc_aiori = {
.name = "S3_EMC",
.create = EMC_Create,
.open = EMC_Open,
.xfer = EMC_Xfer,
.close = EMC_Close,
.delete = S3_Delete,
.set_version = S3_SetVersion,
.fsync = S3_Fsync,
.get_file_size = S3_GetFileSize,
.initialize = S3_init,
.finalize = S3_finalize
};
static void S3_init(void * options){
/* This is supposed to be done before *any* threads are created.
* Could MPI_Init() create threads (or call multi-threaded
* libraries)? We'll assume so. */
AWS4C_CHECK( aws_init() );
}
static void S3_finalize(void * options){
/* done once per program, after exiting all threads.
* NOTE: This fn doesn't return a value that can be checked for success. */
aws_cleanup();
}
static int S3_check_params(IOR_param_t * test){
/* N:1 and N:N */
IOR_offset_t NtoN = test->filePerProc;
IOR_offset_t Nto1 = ! NtoN;
IOR_offset_t s = test->segmentCount;
IOR_offset_t t = test->transferSize;
IOR_offset_t b = test->blockSize;
if (Nto1 && (s != 1) && (b != t)) {
ERR("N:1 (strided) requires xfer-size == block-size");
return 1;
}
return 0;
}
/* modelled on similar macros in iordef.h */
#define CURL_ERR(MSG, CURL_ERRNO, PARAM) \
do { \
fprintf(stdout, "ior ERROR: %s: %s (curl-errno=%d) (%s:%d)\n", \
MSG, curl_easy_strerror(CURL_ERRNO), CURL_ERRNO, \
__FILE__, __LINE__); \
fflush(stdout); \
MPI_Abort((PARAM)->testComm, -1); \
} while (0)
#define CURL_WARN(MSG, CURL_ERRNO) \
do { \
fprintf(stdout, "ior WARNING: %s: %s (curl-errno=%d) (%s:%d)\n", \
MSG, curl_easy_strerror(CURL_ERRNO), CURL_ERRNO, \
__FILE__, __LINE__); \
fflush(stdout); \
} while (0)
/***************************** F U N C T I O N S ******************************/
/* ---------------------------------------------------------------------------
* "Connect" to an S3 object-file-system. We're really just initializing
* libcurl. We need this done before any interactions. It is easy for
* ior_aiori.open/create to assure that we connect, if we haven't already
* done so. However, there's not a simple way to assure that we
* "disconnect" at the end. For now, we'll make a special call at the end
* of ior.c
*
* NOTE: It's okay to call this thing whenever you need to be sure the curl
* handle is initialized.
*
* NOTE: Our custom version of aws4c can be configured so that connections
* are reused, instead of opened and closed on every operation. We
* do configure it that way, but you still need to call these
* connect/disconnect functions, in order to insure that aws4c has
* been configured.
* ---------------------------------------------------------------------------
*/
static void s3_connect( IOR_param_t* param ) {
if (param->verbose >= VERBOSE_2) {
printf("-> s3_connect\n"); /* DEBUGGING */
}
if ( param->curl_flags & IOR_CURL_INIT ) {
if (param->verbose >= VERBOSE_2) {
printf("<- s3_connect [nothing to do]\n"); /* DEBUGGING */
}
return;
}
// --- Done once-only (per rank). Perform all first-time inits.
//
// The aws library requires a config file, as illustrated below. We
// assume that the user running the test has an entry in this file,
// using their login moniker (i.e. `echo $USER`) as the key, as
// suggested in the example:
//
// <user>:<s3_login_id>:<s3_private_key>
//
// This file must not be readable by other than user.
//
// NOTE: These inits could be done in init_IORParam_t(), in ior.c, but
// would require conditional compilation, there.
aws_set_debug(param->verbose >= 4);
aws_read_config(getenv("USER")); // requires ~/.awsAuth
aws_reuse_connections(1);
// initialize IOBufs. These are basically dynamically-extensible
// linked-lists. "growth size" controls the increment of new memory
// allocated, whenever storage is used up.
param->io_buf = aws_iobuf_new();
aws_iobuf_growth_size(param->io_buf, 1024*1024*1);
param->etags = aws_iobuf_new();
aws_iobuf_growth_size(param->etags, 1024*1024*8);
// WARNING: if you have http_proxy set in your environment, you may need
// to override it here. TBD: add a command-line variable to
// allow you to define a proxy.
//
// our hosts are currently 10.140.0.15 - 10.140 0.18
// TBD: Try DNS-round-robin server at vi-lb.ccstar.lanl.gov
// TBD: try HAProxy round-robin at 10.143.0.1
#if 1
// snprintf(buff, BUFF_SIZE, "10.140.0.%d:9020", 15 + (rank % 4));
// s3_set_proxy(buff);
//
// snprintf(buff, BUFF_SIZE, "10.140.0.%d", 15 + (rank % 4));
// s3_set_host(buff);
snprintf(buff, BUFF_SIZE, "10.140.0.%d:9020", 15 + (rank % 4));
s3_set_host(buff);
#else
/*
* If you just want to go to one if the ECS nodes, put that IP
* address in here directly with port 9020.
*
*/
// s3_set_host("10.140.0.15:9020");
/*
* If you want to go to haproxy.ccstar.lanl.gov, this is its IP
* address.
*
*/
// s3_set_proxy("10.143.0.1:80");
// s3_set_host( "10.143.0.1:80");
#endif
// make sure test-bucket exists
s3_set_bucket((char*)bucket_name);
if (rank == 0) {
AWS4C_CHECK( s3_head(param->io_buf, "") );
if ( param->io_buf->code == 404 ) { // "404 Not Found"
printf(" bucket '%s' doesn't exist\n", bucket_name);
AWS4C_CHECK( s3_put(param->io_buf, "") ); /* creates URL as bucket + obj */
AWS4C_CHECK_OK( param->io_buf ); // assure "200 OK"
printf("created bucket '%s'\n", bucket_name);
}
else { // assure "200 OK"
AWS4C_CHECK_OK( param->io_buf );
}
}
MPI_CHECK(MPI_Barrier(param->testComm), "barrier error");
// Maybe allow EMC extensions to S3
s3_enable_EMC_extensions(param->curl_flags & IOR_CURL_S3_EMC_EXT);
// don't perform these inits more than once
param->curl_flags |= IOR_CURL_INIT;
if (param->verbose >= VERBOSE_2) {
printf("<- s3_connect [success]\n");
}
}
static
void
s3_disconnect( IOR_param_t* param ) {
if (param->verbose >= VERBOSE_2) {
printf("-> s3_disconnect\n");
}
// nothing to do here, if using new aws4c ...
if (param->verbose >= VERBOSE_2) {
printf("<- s3_disconnect\n");
}
}
// After finalizing an S3 multi-part-upload, you must reset some things
// before you can use multi-part-upload again. This will also avoid (one
// particular set of) memory-leaks.
void
s3_MPU_reset(IOR_param_t* param) {
aws_iobuf_reset(param->io_buf);
aws_iobuf_reset(param->etags);
param->part_number = 0;
}
/* ---------------------------------------------------------------------------
* direct support for the IOR S3 interface
* ---------------------------------------------------------------------------
*/
/*
* One doesn't "open" an object, in REST semantics. All we really care
* about is whether caller expects the object to have zero-size, when we
* return. If so, we conceptually delete it, then recreate it empty.
*
* ISSUE: If the object is going to receive "appends" (supported in EMC S3
* extensions), the object has to exist before the first append
* operation. On the other hand, there appears to be a bug in the
* EMC implementation, such that if an object ever receives appends,
* and then is deleted, and then recreated, the recreated object will
* always return "500 Server Error" on GET (whether it has been
* appended or not).
*
* Therefore, a safer thing to do here is write zero-length contents,
* instead of deleting.
*
* NOTE: There's also no file-descriptor to return, in REST semantics. On
* the other hand, we keep needing the file *NAME*. Therefore, we
* will return the file-name, and let IOR pass it around to our
* functions, in place of what IOR understands to be a
* file-descriptor.
*
*/
static
void *
S3_Create_Or_Open_internal(char* testFileName,
IOR_param_t* param,
unsigned char createFile,
int multi_part_upload_p ) {
if (param->verbose >= VERBOSE_2) {
printf("-> S3_Create_Or_Open('%s', ,%d, %d)\n",
testFileName, createFile, multi_part_upload_p);
}
/* initialize curl, if needed */
s3_connect( param );
/* Check for unsupported flags */
if ( param->openFlags & IOR_EXCL ) {
fprintf( stdout, "Opening in Exclusive mode is not implemented in S3\n" );
}
if ( param->useO_DIRECT == TRUE ) {
fprintf( stdout, "Direct I/O mode is not implemented in S3\n" );
}
// easier to think
int n_to_n = param->filePerProc;
int n_to_1 = ! n_to_n;
/* check whether object needs reset to zero-length */
int needs_reset = 0;
if (! multi_part_upload_p)
needs_reset = 1; /* so "append" can work */
else if ( param->openFlags & IOR_TRUNC )
needs_reset = 1; /* so "append" can work */
else if (createFile) {
// AWS4C_CHECK( s3_head(param->io_buf, testFileName) );
// if ( ! AWS4C_OK(param->io_buf) )
needs_reset = 1;
}
if ( param->open == WRITE ) {
/* initializations for N:1 or N:N writes using multi-part upload */
if (multi_part_upload_p) {
// For N:N, all ranks do their own MPU open/close. For N:1, only
// rank0 does that. Either way, the response from the server
// includes an "uploadId", which must be used to upload parts to
// the same object.
if ( n_to_n || (rank == 0) ) {
// rank0 handles truncate
if ( needs_reset) {
aws_iobuf_reset(param->io_buf);
AWS4C_CHECK( s3_put(param->io_buf, testFileName) ); /* 0-length write */
AWS4C_CHECK_OK( param->io_buf );
}
// POST request with URL+"?uploads" initiates multi-part upload
snprintf(buff, BUFF_SIZE, "%s?uploads", testFileName);
IOBuf* response = aws_iobuf_new();
AWS4C_CHECK( s3_post2(param->io_buf, buff, NULL, response) );
AWS4C_CHECK_OK( param->io_buf );
// parse XML returned from server, into a tree structure
aws_iobuf_realloc(response);
xmlDocPtr doc = xmlReadMemory(response->first->buf,
response->first->len,
NULL, NULL, 0);
if (doc == NULL)
ERR_SIMPLE("Rank0 Failed to find POST response\n");
// navigate parsed XML-tree to find UploadId
xmlNode* root_element = xmlDocGetRootElement(doc);
const char* upload_id = find_element_named(root_element, (char*)"UploadId");
if (! upload_id)
ERR_SIMPLE("couldn't find 'UploadId' in returned XML\n");
if (param->verbose >= VERBOSE_3)
printf("got UploadId = '%s'\n", upload_id);
const size_t upload_id_len = strlen(upload_id);
if (upload_id_len > MAX_UPLOAD_ID_SIZE) {
snprintf(buff, BUFF_SIZE,
"UploadId length %d exceeds expected max (%d)",
upload_id_len, MAX_UPLOAD_ID_SIZE);
ERR_SIMPLE(buff);
}
// save the UploadId we found
memcpy(param->UploadId, upload_id, upload_id_len);
param->UploadId[upload_id_len] = 0;
// free storage for parsed XML tree
xmlFreeDoc(doc);
aws_iobuf_free(response);
// For N:1, share UploadId across all ranks
if (n_to_1)
MPI_Bcast(param->UploadId, MAX_UPLOAD_ID_SIZE, MPI_BYTE, 0, param->testComm);
}
else
// N:1, and we're not rank0. recv UploadID from Rank 0
MPI_Bcast(param->UploadId, MAX_UPLOAD_ID_SIZE, MPI_BYTE, 0, param->testComm);
}
/* initializations for N:N or N:1 writes using EMC byte-range extensions */
else {
/* maybe reset to zero-length, so "append" can work */
if (needs_reset) {
if (verbose >= VERBOSE_3) {
fprintf( stdout, "rank %d resetting\n",
rank);
}
aws_iobuf_reset(param->io_buf);
AWS4C_CHECK( s3_put(param->io_buf, testFileName) );
AWS4C_CHECK_OK( param->io_buf );
}
}
}
if (param->verbose >= VERBOSE_2) {
printf("<- S3_Create_Or_Open\n");
}
return ((void *) testFileName );
}
static
void *
S3_Create( char *testFileName, IOR_param_t * param ) {
if (param->verbose >= VERBOSE_2) {
printf("-> S3_Create\n");
}
if (param->verbose >= VERBOSE_2) {
printf("<- S3_Create\n");
}
return S3_Create_Or_Open_internal( testFileName, param, TRUE, TRUE );
}
static
void *
EMC_Create( char *testFileName, IOR_param_t * param ) {
if (param->verbose >= VERBOSE_2) {
printf("-> EMC_Create\n");
}
if (param->verbose >= VERBOSE_2) {
printf("<- EMC_Create\n");
}
return S3_Create_Or_Open_internal( testFileName, param, TRUE, FALSE );
}
static
void *
S3_Open( char *testFileName, IOR_param_t * param ) {
if (param->verbose >= VERBOSE_2) {
printf("-> S3_Open\n");
}
if ( param->openFlags & IOR_CREAT ) {
if (param->verbose >= VERBOSE_2) {
printf("<- S3_Open( ... TRUE)\n");
}
return S3_Create_Or_Open_internal( testFileName, param, TRUE, TRUE );
}
else {
if (param->verbose >= VERBOSE_2) {
printf("<- S3_Open( ... FALSE)\n");
}
return S3_Create_Or_Open_internal( testFileName, param, FALSE, TRUE );
}
}
static
void *
EMC_Open( char *testFileName, IOR_param_t * param ) {
if (param->verbose >= VERBOSE_2) {
printf("-> S3_Open\n");
}
if ( param->openFlags & IOR_CREAT ) {
if (param->verbose >= VERBOSE_2) {
printf("<- EMC_Open( ... TRUE)\n");
}
return S3_Create_Or_Open_internal( testFileName, param, TRUE, FALSE );
}
else {
if (param->verbose >= VERBOSE_2) {
printf("<- EMC_Open( ... FALSE)\n");
}
return S3_Create_Or_Open_internal( testFileName, param, FALSE, FALSE );
}
}
/*
* transfer (more) data to an object. <file> is just the obj name.
*
* For N:1, param->offset is understood as offset for a given client to
* write into the "file". This translates to a byte-range in the HTTP
* request. Each write in the N:1 case is treated as a complete "part",
* so there is no such thing as a partial write.
*
* For N:N, when IOR "transfer-size" differs from "block-size", IOR treats
* Xfer as a partial write (i.e. there are multiple calls to XFER, to write
* any one of the "N" objects, as a series of "append" operations). This
* is not supported in S3/REST. Therefore, we depend on an EMC extension,
* in this case. This EMC extension allows appends using a byte-range
* header spec of "Range: bytes=-1-". aws4c now provides
* s3_enable_EMC_extensions(), to allow this behavior. If EMC-extensions
* are not enabled, the aws4c library will generate a run-time error, in
* this case.
*
* Each write-request returns an ETag which is a hash of the data. (The
* ETag could also be computed directly, if we wanted.) We must save the
* etags for later use by S3_close().
*
* WARNING: "Pure" S3 doesn't allow byte-ranges for writes to an object.
* Thus, you also can not append to an object. In the context of IOR,
* this causes objects to have only the size of the most-recent write.
* Thus, If the IOR "transfer-size" is different from the IOR
* "block-size", the files will be smaller than the amount of data
* that was written to them.
*
* EMC does support "append" to an object. In order to allow this,
* you must enable the EMC-extensions in the aws4c library, by calling
* s3_set_emc_compatibility() with a non-zero argument.
*
* NOTE: I don't think REST allows us to read/write an amount other than
* the size we request. Maybe our callback-handlers (above) could
* tell us? For now, this is assuming we only have to send one
* request, to transfer any amount of data. (But see above, re EMC
* support for "append".)
*/
/* In the EMC case, instead of Multi-Part Upload we can use HTTP
* "byte-range" headers to write parts of a single object. This appears to
* have several advantages over the S3 MPU spec:
*
* (a) no need for a special "open" operation, to capture an "UploadID".
* Instead we simply write byte-ranges, and the server-side resolves
* any races, producing a single winner. In the IOR case, there should
* be no races, anyhow.
*
* (b) individual write operations don't have to refer to an ID, or to
* parse and save ETags returned from every write.
*
* (c) no need for a special "close" operation, in which all the saved
* ETags are gathered at a single rank, placed into XML, and shipped to
* the server, to finalize the MPU. That special close appears to
* impose two scaling problems: (1) requires all ETags to be shipped at
* the BW available to a single process, (1) requires either that they
* all fit into memory of a single process, or be written to disk
* (imposes additional BW constraints), or make a more-complex
* interaction with a threaded curl writefunction, to present the
* appearance of a single thread to curl, whilst allowing streaming
* reception of non-local ETags.
*
* (d) no constraints on the number or size of individual parts. (These
* exist in the S3 spec, the EMC's impl of the S3 multi-part upload is
* also free of these constraints.)
*
* Instead, parallel processes can write any number and/or size of updates,
* using a "byte-range" header. After each write returns, that part of the
* global object is visible to any reader. Places that are not updated
* read as zeros.
*/
static
IOR_offset_t
S3_Xfer_internal(int access,
void* file,
IOR_size_t* buffer,
IOR_offset_t length,
IOR_param_t* param,
int multi_part_upload_p ) {
if (param->verbose >= VERBOSE_2) {
printf("-> S3_Xfer(acc:%d, target:%s, buf:0x%llx, len:%llu, 0x%llx)\n",
access, (char*)file, buffer, length, param);
}
char* fname = (char*)file; /* see NOTE above S3_Create_Or_Open() */
size_t remaining = (size_t)length;
char* data_ptr = (char *)buffer;
off_t offset = param->offset;
// easier to think
int n_to_n = param->filePerProc;
int n_to_1 = (! n_to_n);
int segmented = (param->segmentCount == 1);
if (access == WRITE) { /* WRITE */
if (verbose >= VERBOSE_3) {
fprintf( stdout, "rank %d writing length=%lld to offset %lld\n",
rank,
remaining,
param->offset + length - remaining);
}
if (multi_part_upload_p) {
// For N:1, part-numbers must have a global ordering for the
// components of the final object. param->part_number is
// incremented by 1 per write, on each rank. This lets us use it
// to compute a global part-numbering.
//
// In the N:N case, we only need to increment part-numbers within
// each rank.
//
// In the N:1 case, the global order of part-numbers we're writing
// depends on whether wer're writing strided or segmented, in
// other words, how <offset> and <remaining> are actually
// positioning the parts being written. [See discussion at
// S3_Close_internal().]
//
// NOTE: 's3curl.pl --debug' shows StringToSign having partNumber
// first, even if I put uploadId first in the URL. Maybe
// that's what the server will do. GetStringToSign() in
// aws4c is not clever about this, so we spoon-feed args in
// the proper order.
size_t part_number;
if (n_to_1) {
if (segmented) { // segmented
size_t parts_per_rank = param->blockSize / param->transferSize;
part_number = (rank * parts_per_rank) + param->part_number;
}
else // strided
part_number = (param->part_number * param->numTasks) + rank;
}
else
part_number = param->part_number;
++ param->part_number;
// if (verbose >= VERBOSE_3) {
// fprintf( stdout, "rank %d of %d writing (%s,%s) part_number %lld\n",
// rank,
// param->numTasks,
// (n_to_1 ? "N:1" : "N:N"),
// (segmented ? "segmented" : "strided"),
// part_number);
// }
snprintf(buff, BUFF_SIZE,
"%s?partNumber=%d&uploadId=%s",
fname, part_number, param->UploadId);
// For performance, we append <data_ptr> directly into the linked list
// of data in param->io_buf. We are "appending" rather than
// "extending", so the added buffer is seen as written data, rather
// than empty storage.
//
// aws4c parses some header-fields automatically for us (into members
// of the IOBuf). After s3_put2(), we can just read the etag from
// param->io_buf->eTag. The server actually returns literal
// quote-marks, at both ends of the string.
aws_iobuf_reset(param->io_buf);
aws_iobuf_append_static(param->io_buf, data_ptr, remaining);
AWS4C_CHECK( s3_put(param->io_buf, buff) );
AWS4C_CHECK_OK( param->io_buf );
// if (verbose >= VERBOSE_3) {
// printf("rank %d: read ETag = '%s'\n", rank, param->io_buf->eTag);
// if (strlen(param->io_buf->eTag) != ETAG_SIZE+2) { /* quotes at both ends */
// fprintf(stderr, "Rank %d: ERROR: expected ETag to be %d hex digits\n",
// rank, ETAG_SIZE);
// exit(1);
// }
// }
if (verbose >= VERBOSE_3) {
fprintf( stdout, "rank %d of %d (%s,%s) offset %lld, part# %lld --> ETag %s\n",
rank,
param->numTasks,
(n_to_1 ? "N:1" : "N:N"),
(segmented ? "segmented" : "strided"),
offset,
part_number,
param->io_buf->eTag); // incl quote-marks at [0] and [len-1]
}
if (strlen(param->io_buf->eTag) != ETAG_SIZE+2) { /* quotes at both ends */
fprintf(stderr, "Rank %d: ERROR: expected ETag to be %d hex digits\n",
rank, ETAG_SIZE);
exit(1);
}
// save the eTag for later
//
// memcpy(etag, param->io_buf->eTag +1, strlen(param->io_buf->eTag) -2);
// etag[ETAG_SIZE] = 0;
aws_iobuf_append(param->etags,
param->io_buf->eTag +1,
strlen(param->io_buf->eTag) -2);
// DEBUGGING
if (verbose >= VERBOSE_4) {
printf("rank %d: part %d = ETag %s\n", rank, part_number, param->io_buf->eTag);
}
// drop ptrs to <data_ptr>, in param->io_buf
aws_iobuf_reset(param->io_buf);
}
else { // use EMC's byte-range write-support, instead of MPU
// NOTE: You must call 's3_enable_EMC_extensions(1)' for
// byte-ranges to work for writes.
if (n_to_n)
s3_set_byte_range(-1,-1); // EMC header "Range: bytes=-1-" means "append"
else
s3_set_byte_range(offset, remaining);
// For performance, we append <data_ptr> directly into the linked list
// of data in param->io_buf. We are "appending" rather than
// "extending", so the added buffer is seen as written data, rather
// than empty storage.
aws_iobuf_reset(param->io_buf);
aws_iobuf_append_static(param->io_buf, data_ptr, remaining);
AWS4C_CHECK ( s3_put(param->io_buf, file) );
AWS4C_CHECK_OK( param->io_buf );
// drop ptrs to <data_ptr>, in param->io_buf
aws_iobuf_reset(param->io_buf);
}
if ( param->fsyncPerWrite == TRUE ) {
WARN("S3 doesn't support 'fsync'" ); /* does it? */
}
}
else { /* READ or CHECK */
if (verbose >= VERBOSE_3) {
fprintf( stdout, "rank %d reading from offset %lld\n",
rank,
param->offset + length - remaining );
}
// read specific byte-range from the object
// [This is included in the "pure" S3 spec.]
s3_set_byte_range(offset, remaining);
// For performance, we append <data_ptr> directly into the linked
// list of data in param->io_buf. In this case (i.e. reading),
// we're "extending" rather than "appending". That means the
// buffer represents empty storage, which will be filled by the
// libcurl writefunction, invoked via aws4c.
aws_iobuf_reset(param->io_buf);
aws_iobuf_extend_static(param->io_buf, data_ptr, remaining);
AWS4C_CHECK( s3_get(param->io_buf, file) );
if (param->io_buf->code != 206) { /* '206 Partial Content' */
snprintf(buff, BUFF_SIZE,
"Unexpected result (%d, '%s')",
param->io_buf->code, param->io_buf->result);
ERR_SIMPLE(buff);
}
// drop refs to <data_ptr>, in param->io_buf
aws_iobuf_reset(param->io_buf);
}
if (param->verbose >= VERBOSE_2) {
printf("<- S3_Xfer\n");
}
return ( length );
}
static
IOR_offset_t
S3_Xfer(int access,
void* file,
IOR_size_t* buffer,
IOR_offset_t length,
IOR_param_t* param ) {
S3_Xfer_internal(access, file, buffer, length, param, TRUE);
}
static
IOR_offset_t
EMC_Xfer(int access,
void* file,
IOR_size_t* buffer,
IOR_offset_t length,
IOR_param_t* param ) {
S3_Xfer_internal(access, file, buffer, length, param, FALSE);
}
/*
* Does this even mean anything, for HTTP/S3 ?
*
* I believe all interactions with the server are considered complete at
* the time we get a response, e.g. from s3_put(). Therefore, fsync is
* kind of meaningless, for REST/S3.
*
* In future, we could extend our interface so as to allow a non-blocking
* semantics, for example with the libcurl "multi" interface, and/or by
* adding threaded callback handlers to obj_put(). *IF* we do that, *THEN*
* we should revisit 'fsync'.
*
* Another special case is multi-part upload, where many parallel clients
* may be writing to the same "file". (It looks like param->filePerProc
* would be the flag to check, for this.) Maybe when you called 'fsync',
* you meant that you wanted *all* the clients to be complete? That's not
* really what fsync would do. In the N:1 case, this is accomplished by
* S3_Close(). If you really wanted this behavior from S3_Fsync, we could
* have S3_Fsync call S3_close.
*
* As explained above, we may eventually want to consider the following:
*
* (1) thread interaction with any handlers that are doing ongoing
* interactions with the socket, to make sure they have finished all
* actions and gotten responses.
*
* (2) MPI barrier for all clients involved in a multi-part upload.
* Presumably, for IOR, when we are doing N:1, all clients are
* involved in that transfer, so this would amount to a barrier on
* MPI_COMM_WORLD.
*/
static
void
S3_Fsync( void *fd, IOR_param_t * param ) {
if (param->verbose >= VERBOSE_2) {
printf("-> S3_Fsync [no-op]\n");
}
if (param->verbose >= VERBOSE_2) {
printf("<- S3_Fsync\n");
}
}
/*
* It seems the only kind of "close" that ever needs doing for S3 is in the
* case of multi-part upload (i.e. N:1). In this case, all the parties to
* the upload must provide their ETags to a single party (e.g. rank 0 in an
* MPI job). Then the rank doing the closing can generate XML and complete
* the upload.
*
* ISSUE: The S3 spec says that a multi-part upload can have at most 10,000
* parts. Does EMC allow more than this? (NOTE the spec also says
* parts must be at least 5MB, but EMC definitely allows smaller
* parts than that.)
*
* ISSUE: All Etags must be sent from a single rank, in a single
* transaction. If the issue above (regarding 10k Etags) is
* resolved by a discovery that EMC supports more than 10k ETags,
* then, for large-enough files (or small-enough transfer-sizes) an
* N:1 write may generate more ETags than the single closing rank
* can hold in memory. In this case, there are several options,
* outlined
*
*
* See S3_Fsync() for some possible considerations.
*/
static
void
S3_Close_internal( void* fd,
IOR_param_t* param,
int multi_part_upload_p ) {
char* fname = (char*)fd; /* see NOTE above S3_Create_Or_Open() */
// easier to think
int n_to_n = param->filePerProc;
int n_to_1 = (! n_to_n);
int segmented = (param->segmentCount == 1);
if (param->verbose >= VERBOSE_2) {
printf("-> S3_Close('%s', ,%d) %s\n",
fname,
multi_part_upload_p,
((n_to_n) ? "N:N" : ((segmented) ? "N:1(seg)" : "N:1(str)")));
}
if (param->open == WRITE) {
// finalizing Multi-Part Upload (for N:1 or N:N)
if (multi_part_upload_p) {
size_t etag_data_size = param->etags->write_count; /* local ETag data (bytes) */
size_t etags_per_rank = etag_data_size / ETAG_SIZE; /* number of local etags */
// --- create XML containing ETags in an IOBuf for "close" request
IOBuf* xml = NULL;
if (n_to_1) {
// for N:1, gather all Etags at Rank0
MPI_Datatype mpi_size_t;
if (sizeof(size_t) == sizeof(int))
mpi_size_t = MPI_INT;
else if (sizeof(size_t) == sizeof(long))
mpi_size_t = MPI_LONG;
else
mpi_size_t = MPI_LONG_LONG;
// Everybody should have the same number of ETags (?)
size_t etag_count_max = 0; /* highest number on any proc */
MPI_Allreduce(&etags_per_rank, &etag_count_max,
1, mpi_size_t, MPI_MAX, param->testComm);
if (etags_per_rank != etag_count_max) {
printf("Rank %d: etag count mismatch: max:%d, mine:%d\n",
rank, etag_count_max, etags_per_rank);
MPI_Abort(param->testComm, 1);
}
// collect ETag data at Rank0
aws_iobuf_realloc(param->etags); /* force single contiguous buffer */
char* etag_data = param->etags->first->buf; /* per-rank data, contiguous */
if (rank == 0) {
char* etag_ptr;
int i;
int j;
int rnk;
char* etag_vec = (char*)malloc((param->numTasks * etag_data_size) +1);
if (! etag_vec) {
fprintf(stderr, "rank 0 failed to malloc %d bytes\n",
param->numTasks * etag_data_size);
MPI_Abort(param->testComm, 1);
}
MPI_Gather(etag_data, etag_data_size, MPI_BYTE,
etag_vec, etag_data_size, MPI_BYTE, 0, MPI_COMM_WORLD);
// --- debugging: show the gathered etag data
// (This shows the raw concatenated etag-data from each node.)
if (param->verbose >= VERBOSE_4) {
printf("rank 0: gathered %d etags from all ranks:\n", etags_per_rank);
etag_ptr=etag_vec;
for (rnk=0; rnk<param->numTasks; ++rnk) {
printf("\t[%d]: '", rnk);
int ii;
for (ii=0; ii<etag_data_size; ++ii) /* NOT null-terminated! */
printf("%c", etag_ptr[ii]);
printf("'\n");
etag_ptr += etag_data_size;
}
}
// add XML for *all* the parts. The XML must be ordered by
// part-number. Each rank wrote <etags_per_rank> parts,
// locally. At rank0, the etags for each rank are now
// stored as a contiguous block of text, with the blocks
// stored in rank order in etag_vec. In other words, our
// internal rep at rank 0 matches the "segmented" format.
// From this, we must select etags in an order matching how
// they appear in the actual object, and give sequential
// part-numbers to the resulting sequence.
//
// That ordering of parts in the actual written object
// varies according to whether we wrote in the "segmented"
// or "strided" format.
//
// supposing N ranks, and P parts per rank:
//
// segmented:
//
// all parts for a given rank are consecutive.
// rank r writes these parts:
//
// rP, rP+1, ... (r+1)P -1
//
// i.e. rank0 writes parts 0,1,2,3 ... P-1
//
//
// strided:
//
// rank r writes every P-th part, starting with r.
//
// r, P+r, ... (P-1)P + r
//
// i.e. rank0 writes parts 0,P,2P,3P ... (P-1)P
//
//
// NOTE: If we knew ahead of time how many parts each rank was
// going to write, we could assign part-number ranges, per
// rank, and then have nice locality here.
//
// Alternatively, we could have everyone format their own
// XML text and send that, instead of just the tags. This
// would increase the amount of data being sent, but would
// reduce the work for rank0 to format everything.
size_t i_max; // outer-loop
size_t j_max; // inner loop
size_t start_multiplier; // initial offset in collected data
size_t stride; // in etag_vec
if (segmented) { // segmented
i_max = param->numTasks;
j_max = etags_per_rank;
start_multiplier = etag_data_size; /* one rank's-worth of Etag data */
stride = ETAG_SIZE; /* one ETag */
}
else { // strided
i_max = etags_per_rank;
j_max = param->numTasks;
start_multiplier = ETAG_SIZE; /* one ETag */
stride = etag_data_size; /* one rank's-worth of Etag data */
}
xml = aws_iobuf_new();
aws_iobuf_growth_size(xml, 1024 * 8);
// write XML header ...
aws_iobuf_append_str(xml, "<CompleteMultipartUpload>\n");
int part = 0;
for (i=0; i<i_max; ++i) {
etag_ptr=etag_vec + (i * start_multiplier);
for (j=0; j<j_max; ++j) {
// etags were saved as contiguous text. Extract the next one.
char etag[ETAG_SIZE +1];
memcpy(etag, etag_ptr, ETAG_SIZE);
etag[ETAG_SIZE] = 0;
// write XML for next part, with Etag ...
snprintf(buff, BUFF_SIZE,
" <Part>\n"
" <PartNumber>%d</PartNumber>\n"
" <ETag>%s</ETag>\n"
" </Part>\n",
part, etag);
aws_iobuf_append_str(xml, buff);
etag_ptr += stride;
++ part;
}
}
// write XML tail ...
aws_iobuf_append_str(xml, "</CompleteMultipartUpload>\n");
}
else {
MPI_Gather(etag_data, etag_data_size, MPI_BYTE,
NULL, etag_data_size, MPI_BYTE, 0, MPI_COMM_WORLD);
}
}
else { /* N:N */
xml = aws_iobuf_new();
aws_iobuf_growth_size(xml, 1024 * 8);
// write XML header ...
aws_iobuf_append_str(xml, "<CompleteMultipartUpload>\n");
// all parts of our object were written from this rank.
char etag[ETAG_SIZE +1];
int part = 0;
int i;
for (i=0; i<etags_per_rank; ++i) {
// TBD: Instead of reading into etag, then sprintf'ing, then
// copying into xml, we could just read directly into xml
int sz = aws_iobuf_get_raw(param->etags, etag, ETAG_SIZE);
if (sz != ETAG_SIZE) {
snprintf(buff, BUFF_SIZE,
"Read of ETag %d had length %d (not %d)\n",
rank, i, sz, ETAG_SIZE);
ERR_SIMPLE(buff);
}
etag[ETAG_SIZE] = 0;
// write XML for next part, with Etag ...
snprintf(buff, BUFF_SIZE,
" <Part>\n"
" <PartNumber>%d</PartNumber>\n"
" <ETag>%s</ETag>\n"
" </Part>\n",
part, etag);
aws_iobuf_append_str(xml, buff);
++ part;
}
// write XML tail ...
aws_iobuf_append_str(xml, "</CompleteMultipartUpload>\n");
}
// send request to finalize MPU
if (n_to_n || (rank == 0)) {
// DEBUGGING: show the XML we constructed
if (param->verbose >= VERBOSE_3)
debug_iobuf(xml, 1, 1);
// --- POST our XML to the server.
snprintf(buff, BUFF_SIZE,
"%s?uploadId=%s",
fname, param->UploadId);
AWS4C_CHECK ( s3_post(xml, buff) );
AWS4C_CHECK_OK( xml );
aws_iobuf_free(xml);
}
// everybody reset MPU info. Allows another MPU, and frees memory.
s3_MPU_reset(param);
// Everybody meetup, so non-zero ranks won't go trying to stat the
// N:1 file until rank0 has finished the S3 multi-part finalize.
// The object will not appear to exist, until then.
if (n_to_1)
MPI_CHECK(MPI_Barrier(param->testComm), "barrier error");
}
else {
// No finalization is needed, when using EMC's byte-range writing
// support. However, we do need to make sure everyone has
// finished writing, before anyone starts reading.
if (n_to_1) {
MPI_CHECK(MPI_Barrier(param->testComm), "barrier error");
if (param->verbose >= VERBOSE_2)
printf("rank %d: passed barrier\n", rank);
}
}
// After writing, reset the CURL connection, so that caches won't be
// used for reads.
aws_reset_connection();
}
if (param->verbose >= VERBOSE_2) {
printf("<- S3_Close\n");
}
}
static
void
S3_Close( void* fd,
IOR_param_t* param ) {
S3_Close_internal(fd, param, TRUE);
}
static
void
EMC_Close( void* fd,
IOR_param_t* param ) {
S3_Close_internal(fd, param, FALSE);
}
/*
* Delete an object through the S3 interface.
*
* The only reason we separate out EMC version, is because EMC bug means a
* file that was written with appends can't be deleted, recreated, and then
* successfully read.
*/
static
void
S3_Delete( char *testFileName, IOR_param_t * param ) {
if (param->verbose >= VERBOSE_2) {
printf("-> S3_Delete(%s)\n", testFileName);
}
/* maybe initialize curl */
s3_connect( param );
#if 0
// EMC BUG: If file was written with appends, and is deleted,
// Then any future recreation will result in an object that can't be read.
// this
AWS4C_CHECK( s3_delete(param->io_buf, testFileName) );
#else
// just replace with a zero-length object for now
aws_iobuf_reset(param->io_buf);
AWS4C_CHECK ( s3_put(param->io_buf, testFileName) );
#endif
AWS4C_CHECK_OK( param->io_buf );
if (param->verbose >= VERBOSE_2)
printf("<- S3_Delete\n");
}
static
void
EMC_Delete( char *testFileName, IOR_param_t * param ) {
if (param->verbose >= VERBOSE_2) {
printf("-> EMC_Delete(%s)\n", testFileName);
}
/* maybe initialize curl */
s3_connect( param );
#if 0
// EMC BUG: If file was written with appends, and is deleted,
// Then any future recreation will result in an object that can't be read.
// this
AWS4C_CHECK( s3_delete(param->io_buf, testFileName) );
#else
// just replace with a zero-length object for now
aws_iobuf_reset(param->io_buf);
AWS4C_CHECK ( s3_put(param->io_buf, testFileName) );
#endif
AWS4C_CHECK_OK( param->io_buf );
if (param->verbose >= VERBOSE_2)
printf("<- EMC_Delete\n");
}
/*
* HTTP HEAD returns meta-data for a "file".
*
* QUESTION: What should the <size> parameter be, on a HEAD request? Does
* it matter? We don't know how much data they are going to send, but
* obj_get_callback protects us from overruns. Will someone complain if we
* request more data than the header actually takes?
*/
static
IOR_offset_t
S3_GetFileSize(IOR_param_t * param,
MPI_Comm testComm,
char * testFileName) {
if (param->verbose >= VERBOSE_2) {
printf("-> S3_GetFileSize(%s)\n", testFileName);
}
IOR_offset_t aggFileSizeFromStat; /* i.e. "long long int" */
IOR_offset_t tmpMin, tmpMax, tmpSum;
/* make sure curl is connected, and inits are done */
s3_connect( param );
/* send HEAD request. aws4c parses some headers into IOBuf arg. */
AWS4C_CHECK( s3_head(param->io_buf, testFileName) );
if ( ! AWS4C_OK(param->io_buf) ) {
fprintf(stderr, "rank %d: couldn't stat '%s': %s\n",
rank, testFileName, param->io_buf->result);
MPI_Abort(param->testComm, 1);
}
aggFileSizeFromStat = param->io_buf->contentLen;
if (param->verbose >= VERBOSE_2) {
printf("\trank %d: file-size %llu\n", rank, aggFileSizeFromStat);
}
if ( param->filePerProc == TRUE ) {
if (param->verbose >= VERBOSE_2) {
printf("\tall-reduce (1)\n");
}
MPI_CHECK(MPI_Allreduce(&aggFileSizeFromStat,
&tmpSum, /* sum */
1,
MPI_LONG_LONG_INT,
MPI_SUM,
testComm ),
"cannot total data moved" );
aggFileSizeFromStat = tmpSum;
}
else {
if (param->verbose >= VERBOSE_2) {
printf("\tall-reduce (2a)\n");
}
MPI_CHECK(MPI_Allreduce(&aggFileSizeFromStat,
&tmpMin, /* min */
1,
MPI_LONG_LONG_INT,
MPI_MIN,
testComm ),
"cannot total data moved" );
if (param->verbose >= VERBOSE_2) {
printf("\tall-reduce (2b)\n");
}
MPI_CHECK(MPI_Allreduce(&aggFileSizeFromStat,
&tmpMax, /* max */
1,
MPI_LONG_LONG_INT,
MPI_MAX,
testComm ),
"cannot total data moved" );
if ( tmpMin != tmpMax ) {
if ( rank == 0 ) {
WARN( "inconsistent file size by different tasks" );
}
/* incorrect, but now consistent across tasks */
aggFileSizeFromStat = tmpMin;
}
}
if (param->verbose >= VERBOSE_2) {
printf("<- S3_GetFileSize [%llu]\n", aggFileSizeFromStat);
}
return ( aggFileSizeFromStat );
}
Reference in New Issue View Git Blame Copy Permalink