static double ssd_issue_overlapped_ios(ssd_req **reqs, int total, int elem_num, ssd_t *s)
{
double max_cost = 0;
double parunit_op_cost[SSD_MAX_PARUNITS_PER_ELEM];
double parunit_tot_cost[SSD_MAX_PARUNITS_PER_ELEM];
ssd_element_metadata *metadata;
ssd_power_element_stat *power_stat;
int lbn;
int offset;
int i;
int read_cycle = 0;
listnode **parunits;
// all the requests must be of the same type
for (i = 1; i < total; i ++) {
ASSERT(reqs[i]->is_read == reqs[0]->is_read);
}
// is this a set of read requests?
if (reqs[0]->is_read) {
read_cycle = 1;
}
memset(parunit_tot_cost, 0, sizeof(double)*SSD_MAX_PARUNITS_PER_ELEM);
// find the planes to which the reqs are to be issued
metadata = &(s->elements[elem_num].metadata);
power_stat = &(s->elements[elem_num].power_stat);
parunits = ssd_pick_parunits(reqs, total, elem_num, metadata, s);
// repeat until we've served all the requests
while (1) {
double max_op_cost = 0;
double read_xfer_cost = 0.0;
double write_xfer_cost = 0.0;
int active_parunits = 0;
int op_count = 0;
// do we still have any request to service?
for (i = 0; i < SSD_PARUNITS_PER_ELEM(s); i ++) {
if (ll_get_size(parunits[i]) > 0) {
active_parunits ++;
}
}
// no more requests -- get out
if (active_parunits == 0) {
break;
}
// clear this arrays for storing costs
memset(parunit_op_cost, 0, sizeof(double)*SSD_MAX_PARUNITS_PER_ELEM);
// begin a round of serving. we serve one request per
// parallel unit. if an unit has more than one request
// in the list, they have to be serialized.
max_cost = 0;
for (i = 0; i < SSD_PARUNITS_PER_ELEM(s); i ++) {
int size;
size = ll_get_size(parunits[i]);
if (size > 0) {
int apn;
// this parallel unit has a request to serve
ssd_req *r;
listnode *n = ll_get_nth_node(parunits[i], 0);
op_count ++;
ASSERT(op_count <= active_parunits);
// get the request
r = (ssd_req *)n->data;
lbn = ssd_logical_blockno(r->blk, s);
apn = r->blk/s->params.page_size;
offset = (apn/s->params.nelements)%(s->params.pages_per_block-1);
parunit_op_cost[i] = 0;
if (r->is_read) {
int block = metadata->lba_table[lbn];
if(block == -1){
parunit_op_cost[i] = s->params.page_read_latency;
//Micky
ssd_power_flash_calculate(SSD_POWER_FLASH_READ, s->params.page_read_latency, power_stat, s);
}else if(metadata->block_usage[block].log_index == -1){
parunit_op_cost[i] = s->params.page_read_latency;
//Micky
ssd_power_flash_calculate(SSD_POWER_FLASH_READ, s->params.page_read_latency, power_stat, s);
}else{
parunit_op_cost[i] = s->params.page_read_latency;
//Micky
ssd_power_flash_calculate(SSD_POWER_FLASH_READ, s->params.page_read_latency, power_stat, s);
parunit_op_cost[i] += s->params.page_read_latency;
ssd_power_flash_calculate(SSD_POWER_FLASH_READ, s->params.page_read_latency, power_stat, s);
s->spare_read++;
}
//tiel xfer cost
read_xfer_cost += ssd_data_transfer_cost(s,r->count);
} else { //for write
int plane_num = r->plane_num;
// issue the write to the current active page.
// we need to transfer the data across the serial pins for write.
metadata->active_block = metadata->plane_meta[plane_num].active_block;
// check lbn table
if(metadata->lba_table[lbn] == -1 ) {
metadata->lba_table[lbn] = metadata->active_block;
parunit_op_cost[i] = _ssd_write_page_osr(s, metadata, lbn, offset, power_stat);
_ssd_alloc_active_block(plane_num, elem_num, s);
}
else { //if already mapped, check log block
int tmp_block = metadata->lba_table[lbn];
if(metadata->block_usage[tmp_block].page[offset] == -1) {
parunit_op_cost[i] = _ssd_write_page_osr(s, metadata, lbn, offset, power_stat);
}
else {
if (metadata->block_usage[tmp_block].log_index == -1) {
metadata->block_usage[tmp_block].log_index = _ssd_alloc_log_block(plane_num, elem_num, s, tmp_block);
parunit_op_cost[i] = _ssd_write_log_block_osr(s, metadata, lbn, offset, power_stat);
}
else {
if(_last_page_in_log_block(metadata, s, tmp_block)){
int new_block;
parunit_op_cost[i] += ssd_invoke_logblock_cleaning(elem_num, s, lbn);
new_block = metadata->lba_table[lbn];
if(metadata->block_usage[new_block].log_index == -1){
metadata->block_usage[new_block].log_index = _ssd_alloc_log_block(plane_num, elem_num, s, tmp_block);
}
}else{
parunit_op_cost[i] += _ssd_write_log_block_osr(s, metadata, lbn, offset, power_stat);
}
}
}
}
write_xfer_cost += ssd_data_transfer_cost(s,r->count);
}
ASSERT(r->count <= s->params.page_size);
// calc the cost: the access time should be something like this
// for read
if (read_cycle) {
if (SSD_PARUNITS_PER_ELEM(s) > 4) {
printf("modify acc time here ...\n");
ASSERT(0);
}
if (op_count == 1) {
r->acctime = parunit_op_cost[i] + read_xfer_cost;
r->schtime = parunit_tot_cost[i] + r->acctime;
} else {
r->acctime = ssd_data_transfer_cost(s,r->count);
r->schtime = parunit_tot_cost[i] + read_xfer_cost + parunit_op_cost[i];
}
} else {
// for write
r->acctime = parunit_op_cost[i];
r->schtime = parunit_tot_cost[i] + write_xfer_cost + r->acctime;
}
// find the maximum cost for this round of operations
if (max_cost < r->schtime) {
max_cost = r->schtime;
}
// release the node from the linked list
ll_release_node(parunits[i], n);
}
}
ssd_power_flash_calculate(SSD_POWER_FLASH_BUS_DATA_TRANSFER, read_xfer_cost, power_stat, s);
ssd_power_flash_calculate(SSD_POWER_FLASH_BUS_DATA_TRANSFER, write_xfer_cost, power_stat, s);
// we can start the next round of operations only after all
// the operations in the first round are over because we're
// limited by the one set of pins to all the parunits
for (i = 0; i < SSD_PARUNITS_PER_ELEM(s); i ++) {
parunit_tot_cost[i] = max_cost;
}
}
for (i = 0; i < SSD_PARUNITS_PER_ELEM(s); i ++) {
ll_release(parunits[i]);
}
free(parunits);
power_stat->acc_time += max_cost;
return max_cost;
}
static double ssd_issue_overlapped_ios(ssd_req **reqs, int total, int elem_num, ssd_t *s)
{
double max_cost = 0;
double parunit_op_cost[SSD_MAX_PARUNITS_PER_ELEM];
double parunit_tot_cost[SSD_MAX_PARUNITS_PER_ELEM];
ssd_element_metadata *metadata;
int lpn;
int i;
int read_cycle = 0;
listnode **parunits;
// all the requests must be of the same type
for (i = 1; i < total; i ++) {
ASSERT(reqs[i]->is_read == reqs[0]->is_read);
}
// is this a set of read requests?
if (reqs[0]->is_read) {
read_cycle = 1;
}
memset(parunit_tot_cost, 0, sizeof(double)*SSD_MAX_PARUNITS_PER_ELEM);
// find the planes to which the reqs are to be issued
metadata = &(s->elements[elem_num].metadata);
parunits = ssd_pick_parunits(reqs, total, elem_num, metadata, s);
// repeat until we've served all the requests
while (1) {
//double tot_xfer_cost = 0;
double max_op_cost = 0;
int active_parunits = 0;
int op_count = 0;
// do we still have any request to service?
for (i = 0; i < SSD_PARUNITS_PER_ELEM(s); i ++) {
if (ll_get_size(parunits[i]) > 0) {
active_parunits ++;
}
}
// no more requests -- get out
if (active_parunits == 0) {
break;
}
// clear this arrays for storing costs
memset(parunit_op_cost, 0, sizeof(double)*SSD_MAX_PARUNITS_PER_ELEM);
// begin a round of serving. we serve one request per
// parallel unit. if an unit has more than one request
// in the list, they have to be serialized.
max_cost = 0;
for (i = 0; i < SSD_PARUNITS_PER_ELEM(s); i ++) {
int size;
size = ll_get_size(parunits[i]);
if (size > 0) {
// this parallel unit has a request to serve
ssd_req *r;
listnode *n = ll_get_nth_node(parunits[i], 0);
op_count ++;
ASSERT(op_count <= active_parunits);
// get the request
r = (ssd_req *)n->data;
lpn = ssd_logical_pageno(r->blk, s);
if (r->is_read) {
parunit_op_cost[i] = s->params.page_read_latency;
} else {
int plane_num = r->plane_num;
// if this is the last page on the block, allocate a new block
if (ssd_last_page_in_block(metadata->plane_meta[plane_num].active_page, s)) {
_ssd_alloc_active_block(plane_num, elem_num, s);
}
// issue the write to the current active page.
// we need to transfer the data across the serial pins for write.
metadata->active_page = metadata->plane_meta[plane_num].active_page;
//printf("elem %d plane %d ", elem_num, plane_num);
parunit_op_cost[i] = _ssd_write_page_osr(s, metadata, lpn);
}
ASSERT(r->count <= s->params.page_size);
// calc the cost: the access time should be something like this
// for read
if (read_cycle) {
if (SSD_PARUNITS_PER_ELEM(s) > 4) {
printf("modify acc time here ...\n");
ASSERT(0);
}
if (op_count == 1) {
r->acctime = parunit_op_cost[i] + ssd_data_transfer_cost(s,s->params.page_size);
r->schtime = parunit_tot_cost[i] + (op_count-1)*ssd_data_transfer_cost(s,s->params.page_size) + r->acctime;
} else {
r->acctime = ssd_data_transfer_cost(s,s->params.page_size);
r->schtime = parunit_tot_cost[i] + op_count*ssd_data_transfer_cost(s,s->params.page_size) + parunit_op_cost[i];
}
} else {
// for write
r->acctime = parunit_op_cost[i] + ssd_data_transfer_cost(s,s->params.page_size);
r->schtime = parunit_tot_cost[i] + (op_count-1)*ssd_data_transfer_cost(s,s->params.page_size) + r->acctime;
}
// find the maximum cost for this round of operations
if (max_cost < r->schtime) {
max_cost = r->schtime;
}
// release the node from the linked list
ll_release_node(parunits[i], n);
}
}
// we can start the next round of operations only after all
// the operations in the first round are over because we're
// limited by the one set of pins to all the parunits
for (i = 0; i < SSD_PARUNITS_PER_ELEM(s); i ++) {
parunit_tot_cost[i] = max_cost;
}
}
for (i = 0; i < SSD_PARUNITS_PER_ELEM(s); i ++) {
ll_release(parunits[i]);
}
free(parunits);
return max_cost;
}