ligang/vpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

dpdk: apply dual loop unrolling in DPDK TX

Browse Source 
Too many prefetches within loop unrollings induce bottleneck and
performance degradation on some CPUs which have less cache line fill
buffers, e.g, Arm Cortex-A72.
Apply dual loop unrolling and tune prefetches manually to remove
hot-spot with prefetch instructions, to get throughput improvement.
It brings about 1% throughput improvement and saves 8% clocks with
the target node on Cortex-A72.

Type: feature

Change-Id: If3a64a04a77e90cd0240bc4d1186dbb09dac7df0
Signed-off-by: Lijian Zhang <Lijian.Zhang@arm.com>
This commit is contained in:
Lijian.Zhang
2019-07-11 16:44:22 +08:00
committed by Damjan Marion
parent 8a1dea4ce6
commit fe2523d1a4
1 changed files with 57 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
src/plugins/dpdk/device/device.c
View File

@ -289,6 +289,7 @@ VNET_DEVICE_CLASS_TX_FN (dpdk_device_class) (vlib_main_t * vm,
n_left = n_packets;
mb = ptd->mbufs;
#if (CLIB_N_PREFETCHES >= 8)
while (n_left >= 8)
{
u32 or_flags;
@ -353,6 +354,62 @@ VNET_DEVICE_CLASS_TX_FN (dpdk_device_class) (vlib_main_t * vm,
mb += 4;
n_left -= 4;
}
#elif (CLIB_N_PREFETCHES >= 4)
while (n_left >= 4)
{
vlib_buffer_t *b2, *b3;
u32 or_flags;
CLIB_PREFETCH (mb[2], CLIB_CACHE_LINE_BYTES, STORE);
CLIB_PREFETCH (mb[3], CLIB_CACHE_LINE_BYTES, STORE);
b2 = vlib_buffer_from_rte_mbuf (mb[2]);
CLIB_PREFETCH (b2, CLIB_CACHE_LINE_BYTES, LOAD);
b3 = vlib_buffer_from_rte_mbuf (mb[3]);
CLIB_PREFETCH (b3, CLIB_CACHE_LINE_BYTES, LOAD);
b[0] = vlib_buffer_from_rte_mbuf (mb[0]);
b[1] = vlib_buffer_from_rte_mbuf (mb[1]);
or_flags = b[0]->flags | b[1]->flags;
all_or_flags |= or_flags;
VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[0]);
VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b[1]);
if (or_flags & VLIB_BUFFER_NEXT_PRESENT)
{
dpdk_validate_rte_mbuf (vm, b[0], 1);
dpdk_validate_rte_mbuf (vm, b[1], 1);
}
else
{
dpdk_validate_rte_mbuf (vm, b[0], 0);
dpdk_validate_rte_mbuf (vm, b[1], 0);
}
if (PREDICT_FALSE ((xd->flags & DPDK_DEVICE_FLAG_TX_OFFLOAD) &&
(or_flags &
(VNET_BUFFER_F_OFFLOAD_TCP_CKSUM
| VNET_BUFFER_F_OFFLOAD_IP_CKSUM
| VNET_BUFFER_F_OFFLOAD_UDP_CKSUM))))
{
dpdk_buffer_tx_offload (xd, b[0], mb[0]);
dpdk_buffer_tx_offload (xd, b[1], mb[1]);
}
if (PREDICT_FALSE (node->flags & VLIB_NODE_FLAG_TRACE))
{
if (b[0]->flags & VLIB_BUFFER_IS_TRACED)
dpdk_tx_trace_buffer (dm, node, xd, queue_id, b[0]);
if (b[1]->flags & VLIB_BUFFER_IS_TRACED)
dpdk_tx_trace_buffer (dm, node, xd, queue_id, b[1]);
}
mb += 2;
n_left -= 2;
}
#endif
while (n_left > 0)
{
b[0] = vlib_buffer_from_rte_mbuf (mb[0]);