1
1
Fork 0
mirror of https://github.com/pbatard/rufus.git synced 2024-08-14 23:57:05 +00:00

[checksum] more load balancing optimizations

* SetChecksumAffinity() now spreads the affinity evenly between cores
* Also increase the read buffer size to help with performance
* At this stage, the only limiting factor regarding performance seems
  to be the speed of the SHA-256 algorithm...
This commit is contained in:
Pete Batard 2016-03-01 17:13:37 +00:00
parent e6d3653cac
commit e1c7c9670b
5 changed files with 77 additions and 46 deletions

View file

@ -60,7 +60,7 @@
#undef BIG_ENDIAN_HOST
#define BUFFER_SIZE 4096
#define BUFFER_SIZE (64*KB)
#define WAIT_TIME 5000
/* Globals */
@ -704,6 +704,13 @@ static void md5_final(SUM_CONTEXT *ctx)
#undef X
}
typedef void sum_init_t(SUM_CONTEXT *ctx);
typedef void sum_write_t(SUM_CONTEXT *ctx, const unsigned char *buf, size_t len);
typedef void sum_final_t(SUM_CONTEXT *ctx);
sum_init_t *sum_init[NUM_CHECKSUMS] = { md5_init, sha1_init , sha256_init };
sum_write_t *sum_write[NUM_CHECKSUMS] = { md5_write, sha1_write , sha256_write };
sum_final_t *sum_final[NUM_CHECKSUMS] = { md5_final, sha1_final , sha256_final };
/*
* Checksum dialog callback
*/
@ -771,53 +778,47 @@ INT_PTR CALLBACK ChecksumCallback(HWND hDlg, UINT message, WPARAM wParam, LPARAM
return (INT_PTR)FALSE;
}
typedef void sum_init_t(SUM_CONTEXT *ctx);
typedef void sum_write_t(SUM_CONTEXT *ctx, const unsigned char *buf, size_t len);
typedef void sum_final_t(SUM_CONTEXT *ctx);
sum_init_t *sum_init[NUM_CHECKSUMS] = { md5_init, sha1_init , sha256_init };
sum_write_t *sum_write[NUM_CHECKSUMS] = { md5_write, sha1_write , sha256_write };
sum_final_t *sum_final[NUM_CHECKSUMS] = { md5_final, sha1_final , sha256_final };
/*
* We want the maximum speed we can get out of the checksum computation,
* so, if we have a multiprocessor/multithreaded machine, we'll assign of
* each of the individual checksum threads to a specific virtual core, and
* assign the read thread to one of the remainder virtual cores.
* so, if we have a multiprocessor/multithreaded machine, we try to assign
* each of the individual checksum threads to a different core.
* To do just that, we need the following function call.
* Oh, and BOY is this thing sensitive to whether the first sum affinity
* is on an even or odd virtual core!
*/
BOOL SetChecksumAffinity(CHECKSUM_AFFINITY* checksum_affinity)
extern BOOL usb_debug; // For uuprintf
BOOL SetChecksumAffinity(DWORD_PTR* thread_affinity)
{
int i, pc;
int i, j, pc;
DWORD_PTR affinity, dummy;
memset(checksum_affinity, 0, sizeof(CHECKSUM_AFFINITY));
memset(thread_affinity, 0, 4 * sizeof(DWORD_PTR));
if (!GetProcessAffinityMask(GetCurrentProcess(), &affinity, &dummy))
return FALSE;
uuprintf("\r\nChecksum affinities:");
uuprintf("global:\t%s", printbitslz(affinity));
// If we don't have enough virtual cores to evenly spread our load forget it
pc = popcnt64(affinity);
if (pc < NUM_CHECKSUMS + 1)
return FALSE;
// We'll use the NUM_CHECKSUMS least significant set bits in our mask for
// the individual checksum threads, and the remainder for the read thread.
// From an empirical perspective, this looks like the best "one-size-fits-all"
// to spread the load.
checksum_affinity->read_thread = affinity;
// Spread the affinity as evenly as we can
thread_affinity[NUM_CHECKSUMS] = affinity;
for (i = 0; i < NUM_CHECKSUMS; i++) {
checksum_affinity->sum_thread[i] = affinity & (-1LL * affinity);
affinity ^= checksum_affinity->sum_thread[i];
checksum_affinity->read_thread ^= checksum_affinity->sum_thread[i];
for (j = 0; j < pc / (NUM_CHECKSUMS + 1); j++) {
thread_affinity[i] |= affinity & (-1LL * affinity);
affinity ^= affinity & (-1LL * affinity);
}
uuprintf("sum%d:\t%s", i, printbitslz(thread_affinity[i]));
thread_affinity[NUM_CHECKSUMS] ^= thread_affinity[i];
}
uuprintf("sum%d:\t%s", i, printbitslz(thread_affinity[i]));
return TRUE;
}
// Individual thread that computes one of MD5, SHA1 or SHA256 in parallel
DWORD WINAPI IndividualSumThread(void* param)
{
SUM_CONTEXT sum_ctx;
SUM_CONTEXT sum_ctx = { 0 }; // There's a memset in sum_init, but static analyzers still bug us
int i = (int)(uintptr_t)param, j;
sum_init[i](&sum_ctx);
@ -850,24 +851,28 @@ error:
DWORD WINAPI SumThread(void* param)
{
CHECKSUM_AFFINITY* checksum_affinity = (CHECKSUM_AFFINITY*)param;
DWORD_PTR* thread_affinity = (DWORD_PTR*)param;
HANDLE sum_thread[NUM_CHECKSUMS] = { NULL, NULL, NULL };
HANDLE h = INVALID_HANDLE_VALUE;
uint64_t rb, LastRefresh = 0;
int i, _bufnum, r = -1;
float format_percent = 0.0f;
if ((image_path == NULL) || (checksum_affinity == NULL))
if ((image_path == NULL) || (thread_affinity == NULL))
goto out;
uprintf("\r\nComputing checksum for '%s'...", image_path);
if (checksum_affinity->read_thread != 0)
SetThreadAffinityMask(GetCurrentThread(), checksum_affinity->read_thread);
if (thread_affinity[0] != 0)
// Use the first affinity mask, as our read thread is the least
// CPU intensive (mostly waits on disk I/O or on the other threads)
// whereas the OS is likely to requisition the first Core, which
// is usually in this first mask, for other tasks.
SetThreadAffinityMask(GetCurrentThread(), thread_affinity[0]);
for (i = 0; i < NUM_CHECKSUMS; i++) {
// NB: Can't use a single manual-reset event for data_ready as we
// wouldn't be able to ensure the event is reset before the threa
// wouldn't be able to ensure the event is reset before the thread
// gets into its next wait loop
data_ready[i] = CreateEvent(NULL, FALSE, FALSE, NULL);
thread_ready[i] = CreateEvent(NULL, FALSE, FALSE, NULL);
@ -880,8 +885,8 @@ DWORD WINAPI SumThread(void* param)
uprintf("Unable to start checksum thread #%d", i);
goto out;
}
if (checksum_affinity->sum_thread[i] != 0)
SetThreadAffinityMask(sum_thread[i], checksum_affinity->sum_thread[i]);
if (thread_affinity[i+1] != 0)
SetThreadAffinityMask(sum_thread[i], thread_affinity[i+1]);
}
h = CreateFileU(image_path, GENERIC_READ, FILE_SHARE_READ, NULL,

View file

@ -2045,7 +2045,7 @@ static INT_PTR CALLBACK MainCallback(HWND hDlg, UINT message, WPARAM wParam, LPA
static ULONG ulRegister = 0;
static LPITEMIDLIST pidlDesktop = NULL;
static MY_SHChangeNotifyEntry NotifyEntry;
static CHECKSUM_AFFINITY checksum_affinity;
static DWORD_PTR sumthread_affinity[4];
DRAWITEMSTRUCT* pDI;
HDROP droppedFileInfo;
POINT Point;
@ -2530,8 +2530,8 @@ static INT_PTR CALLBACK MainCallback(HWND hDlg, UINT message, WPARAM wParam, LPA
// Disable all controls except cancel
EnableControls(FALSE);
InitProgress(FALSE);
SetChecksumAffinity(&checksum_affinity);
format_thid = CreateThread(NULL, 0, SumThread, (LPVOID)&checksum_affinity, 0, NULL);
SetChecksumAffinity(sumthread_affinity);
format_thid = CreateThread(NULL, 0, SumThread, (LPVOID)sumthread_affinity, 0, NULL);
if (format_thid != NULL) {
PrintInfo(0, -1);
timer = 0;

View file

@ -291,11 +291,6 @@ typedef struct {
char* path;
} VHD_SAVE;
typedef struct {
DWORD_PTR read_thread;
DWORD_PTR sum_thread[NUM_CHECKSUMS];
} CHECKSUM_AFFINITY;
/*
* Structure and macros used for the extensions specification of FileDialog()
* You can use:
@ -446,7 +441,10 @@ extern LONG ValidateSignature(HWND hDlg, const char* path);
extern BOOL IsFontAvailable(const char* font_name);
extern BOOL WriteFileWithRetry(HANDLE hFile, LPCVOID lpBuffer, DWORD nNumberOfBytesToWrite,
LPDWORD lpNumberOfBytesWritten, DWORD nNumRetries);
extern BOOL SetChecksumAffinity(CHECKSUM_AFFINITY* checksum_affinity);
extern BOOL SetChecksumAffinity(DWORD_PTR* thread_affinity);
#define printbits(x) _printbits(sizeof(x), &x, 0)
#define printbitslz(x) _printbits(sizeof(x), &x, 1)
extern char* _printbits(size_t const size, void const * const ptr, int leading_zeroes);
DWORD WINAPI FormatThread(void* param);
DWORD WINAPI SaveImageThread(void* param);

View file

@ -33,7 +33,7 @@ LANGUAGE LANG_NEUTRAL, SUBLANG_NEUTRAL
IDD_DIALOG DIALOGEX 12, 12, 242, 376
STYLE DS_SETFONT | DS_MODALFRAME | DS_CENTER | WS_MINIMIZEBOX | WS_POPUP | WS_CAPTION | WS_SYSMENU
EXSTYLE WS_EX_ACCEPTFILES
CAPTION "Rufus 2.8.871"
CAPTION "Rufus 2.8.872"
FONT 8, "Segoe UI Symbol", 400, 0, 0x0
BEGIN
LTEXT "Device",IDS_DEVICE_TXT,9,6,200,8
@ -320,8 +320,8 @@ END
//
VS_VERSION_INFO VERSIONINFO
FILEVERSION 2,8,871,0
PRODUCTVERSION 2,8,871,0
FILEVERSION 2,8,872,0
PRODUCTVERSION 2,8,872,0
FILEFLAGSMASK 0x3fL
#ifdef _DEBUG
FILEFLAGS 0x1L
@ -338,13 +338,13 @@ BEGIN
BEGIN
VALUE "CompanyName", "Akeo Consulting (http://akeo.ie)"
VALUE "FileDescription", "Rufus"
VALUE "FileVersion", "2.8.871"
VALUE "FileVersion", "2.8.872"
VALUE "InternalName", "Rufus"
VALUE "LegalCopyright", "© 2011-2016 Pete Batard (GPL v3)"
VALUE "LegalTrademarks", "http://www.gnu.org/copyleft/gpl.html"
VALUE "OriginalFilename", "rufus.exe"
VALUE "ProductName", "Rufus"
VALUE "ProductVersion", "2.8.871"
VALUE "ProductVersion", "2.8.872"
END
END
BLOCK "VarFileInfo"

View file

@ -74,6 +74,34 @@ void _uprintf(const char *format, ...)
}
#endif
// Prints a bitstring of a number of any size, with or without leading zeroes.
// See also the printbits() and printbitslz() helper macros in rufus.h
char *_printbits(size_t const size, void const * const ptr, int leading_zeroes)
{
// sizeof(uintmax_t) so that we have enough space to store whatever is thrown at us
static char str[sizeof(uintmax_t) * 8 + 3];
size_t i;
uint8_t* b = (uint8_t*)ptr;
uintmax_t mask, lzmask = 0, val = 0;
// Little endian, the SCOURGE of any rational computing
for (i = 0; i < size; i++)
val |= ((uintmax_t)b[i]) << (8 * i);
str[0] = '0';
str[1] = 'b';
if (leading_zeroes)
lzmask = 1ULL << (size * 8 - 1);
for (i = 2, mask = 1ULL << (sizeof(uintmax_t) * 8 - 1); mask != 0; mask >>= 1) {
if ((i > 2) || (lzmask & mask))
str[i++] = (val & mask) ? '1' : '0';
else if (val & mask)
str[i++] = '1';
}
str[i] = '\0';
return str;
}
void DumpBufferHex(void *buf, size_t size)
{
unsigned char* buffer = (unsigned char*)buf;