Hello Roland,
Here is the patch to tune IPoIB QP send/recv queue size. Please reivew. The attachement file is for you to apply the patch.
This patch includes:
a. these two parameters are module parameters
b. they are saved in priv_dev
c. the QP size max value=4k, min value=64
d. QP size check to be power of 2 because of tx_ring
e. these QP sizes are logged in /var/log/messages.
f. modify TX_RING/RX_RING_SIZE to SENDQ/RECVQ_SIZE. the ring concept here is only to have a place to save pointers not an actual ring.
This patch will address packets retransmission/timeout issues on large cluster environment. Also the performance has dramatically improved by tuning these parameters. It's about 7 times better throughput than default value according to our MPI test results on cluster.
Signed-off-by: Shirley Ma <[EMAIL PROTECTED]>
diff -urN infiniband/ulp/ipoib/ipoib.h infiniband-queue/ulp/ipoib/ipoib.h
--- infiniband/ulp/ipoib/ipoib.h 2006-03-26 11:57:15.000000000 -0800
+++ infiniband-queue/ulp/ipoib/ipoib.h 2006-03-31 08:46:34.171748048 -0800
@@ -66,8 +66,8 @@
IPOIB_ENCAP_LEN = 4,
- IPOIB_RX_RING_SIZE = 128,
- IPOIB_TX_RING_SIZE = 64,
+ IPOIB_SENDQ_SIZE = 64,
+ IPOIB_RECVQ_SIZE = 128,
IPOIB_NUM_WC = 4,
@@ -186,6 +186,8 @@
struct dentry *mcg_dentry;
struct dentry *path_dentry;
#endif
+ int sendq_size;
+ int recvq_size;
};
struct ipoib_ah {
@@ -338,6 +340,8 @@
#define ipoib_warn(priv, format, arg...) \
ipoib_printk(KERN_WARNING, priv, format , ## arg)
+extern int ipoib_sendq_size;
+extern int ipoib_recvq_size;
#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
extern int ipoib_debug_level;
diff -urN infiniband/ulp/ipoib/ipoib_ib.c infiniband-queue/ulp/ipoib/ipoib_ib.c
--- infiniband/ulp/ipoib/ipoib_ib.c 2006-03-26 11:57:15.000000000 -0800
+++ infiniband-queue/ulp/ipoib/ipoib_ib.c 2006-03-31 08:46:34.227739536 -0800
@@ -161,7 +161,7 @@
struct ipoib_dev_priv *priv = netdev_priv(dev);
int i;
- for (i = 0; i < IPOIB_RX_RING_SIZE; ++i) {
+ for (i = 0; i < priv->recvq_size; ++i) {
if (ipoib_alloc_rx_skb(dev, i)) {
ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
return -ENOMEM;
@@ -187,7 +187,7 @@
if (wr_id & IPOIB_OP_RECV) {
wr_id &= ~IPOIB_OP_RECV;
- if (wr_id < IPOIB_RX_RING_SIZE) {
+ if (wr_id < priv->recvq_size) {
struct sk_buff *skb = priv->rx_ring[wr_id].skb;
dma_addr_t addr = priv->rx_ring[wr_id].mapping;
@@ -252,9 +252,9 @@
struct ipoib_tx_buf *tx_req;
unsigned long flags;
- if (wr_id >= IPOIB_TX_RING_SIZE) {
+ if (wr_id >= priv->sendq_size) {
ipoib_warn(priv, "completion event with wrid %d (> %d)\n",
- wr_id, IPOIB_TX_RING_SIZE);
+ wr_id, priv->sendq_size);
return;
}
@@ -275,7 +275,7 @@
spin_lock_irqsave(&priv->tx_lock, flags);
++priv->tx_tail;
if (netif_queue_stopped(dev) &&
- priv->tx_head - priv->tx_tail <= IPOIB_TX_RING_SIZE / 2)
+ priv->tx_head - priv->tx_tail <= priv->sendq_size / 2)
netif_wake_queue(dev);
spin_unlock_irqrestore(&priv->tx_lock, flags);
@@ -344,13 +344,13 @@
* means we have to make sure everything is properly recorded and
* our state is consistent before we call post_send().
*/
- tx_req = &priv->tx_ring[priv->tx_head & (IPOIB_TX_RING_SIZE - 1)];
+ tx_req = &priv->tx_ring[priv->tx_head & (priv->sendq_size - 1)];
tx_req->skb = skb;
addr = dma_map_single(priv->ca->dma_device, skb->data, skb->len,
DMA_TO_DEVICE);
pci_unmap_addr_set(tx_req, mapping, addr);
- if (unlikely(post_send(priv, priv->tx_head & (IPOIB_TX_RING_SIZE - 1),
+ if (unlikely(post_send(priv, priv->tx_head & (priv->sendq_size - 1),
address->ah, qpn, addr, skb->len))) {
ipoib_warn(priv, "post_send failed\n");
++priv->stats.tx_errors;
@@ -363,7 +363,7 @@
address->last_send = priv->tx_head;
++priv->tx_head;
- if (priv->tx_head - priv->tx_tail == IPOIB_TX_RING_SIZE) {
+ if (priv->tx_head - priv->tx_tail == priv->sendq_size) {
ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
netif_stop_queue(dev);
}
@@ -488,7 +488,7 @@
int pending = 0;
int i;
- for (i = 0; i < IPOIB_RX_RING_SIZE; ++i)
+ for (i = 0; i < priv->recvq_size; ++i)
if (priv->rx_ring[i].skb)
++pending;
@@ -527,7 +527,7 @@
*/
while ((int) priv->tx_tail - (int) priv->tx_head < 0) {
tx_req = &priv->tx_ring[priv->tx_tail &
- (IPOIB_TX_RING_SIZE - 1)];
+ (priv->sendq_size - 1)];
dma_unmap_single(priv->ca->dma_device,
pci_unmap_addr(tx_req, mapping),
tx_req->skb->len,
@@ -536,7 +536,7 @@
++priv->tx_tail;
}
- for (i = 0; i < IPOIB_RX_RING_SIZE; ++i)
+ for (i = 0; i < priv->recvq_size; ++i)
if (priv->rx_ring[i].skb) {
dma_unmap_single(priv->ca->dma_device,
pci_unmap_addr(&priv->rx_ring[i],
diff -urN infiniband/ulp/ipoib/ipoib_main.c infiniband-queue/ulp/ipoib/ipoib_main.c
--- infiniband/ulp/ipoib/ipoib_main.c 2006-03-28 19:20:21.000000000 -0800
+++ infiniband-queue/ulp/ipoib/ipoib_main.c 2006-03-31 09:15:06.345458080 -0800
@@ -53,6 +53,17 @@
MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
MODULE_LICENSE("Dual BSD/GPL");
+#define IPOIB_MAX_QUEUE_SIZE 4096 /* max is 4k */
+#define IPOIB_MIN_QUEUE_SIZE 64 /* min is 64 */
+
+int ipoib_sendq_size = IPOIB_SENDQ_SIZE;
+int ipoib_recvq_size = IPOIB_RECVQ_SIZE;
+
+module_param_named(sendq_size, ipoib_sendq_size, int, 0444);
+MODULE_PARM_DESC(sendq_size, "Number of wqe in send queue");
+module_param_named(recvq_size, ipoib_recvq_size, int, 0444);
+MODULE_PARM_DESC(recvq_size, "Number of wqe in receive queue");
+
#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
int ipoib_debug_level;
@@ -837,27 +848,61 @@
return 0;
}
+static int expsize(int size)
+{
+ int expsize_t = 1;
+ int j = 1;
+ while (size / 2 >= expsize_t) {
+ expsize_t = 1 << ++j;
+ }
+ return expsize_t;
+}
+
int ipoib_dev_init(struct net_device *dev, struct ib_device *ca, int port)
{
struct ipoib_dev_priv *priv = netdev_priv(dev);
/* Allocate RX/TX "rings" to hold queued skbs */
- priv->rx_ring = kzalloc(IPOIB_RX_RING_SIZE * sizeof (struct ipoib_rx_buf),
+ if (ipoib_recvq_size > IPOIB_MAX_QUEUE_SIZE) {
+ ipoib_recvq_size = IPOIB_MAX_QUEUE_SIZE;
+ printk(KERN_WARNING "%s: ipoib_recvq_size is too big, use max %d instead\n", ca->name, IPOIB_MAX_QUEUE_SIZE);
+ }
+ if (ipoib_recvq_size < IPOIB_MIN_QUEUE_SIZE) {
+ ipoib_recvq_size = IPOIB_MIN_QUEUE_SIZE;
+ printk(KERN_WARNING "%s: ipoib_recvq_size is too small, use min %d instead\n", ca->name, IPOIB_MIN_QUEUE_SIZE);
+ }
+ priv->recvq_size = expsize(ipoib_recvq_size);
+ priv->rx_ring = kzalloc(priv->recvq_size * sizeof (struct ipoib_rx_buf),
GFP_KERNEL);
if (!priv->rx_ring) {
printk(KERN_WARNING "%s: failed to allocate RX ring (%d entries)\n",
- ca->name, IPOIB_RX_RING_SIZE);
+ ca->name, priv->sendq_size);
goto out;
}
+ printk(KERN_INFO "%s: RX_RING_SIZE is set to %d entries\n",
+ ca->name, priv->recvq_size);
+
+ if (ipoib_sendq_size > IPOIB_MAX_QUEUE_SIZE) {
+ ipoib_sendq_size = IPOIB_MAX_QUEUE_SIZE;
+ printk(KERN_WARNING "%s: ipoib_sendq_size is too big, use max %d instead\n", ca->name, IPOIB_MAX_QUEUE_SIZE);
+ }
+ if (ipoib_sendq_size < IPOIB_MIN_QUEUE_SIZE) {
+ ipoib_sendq_size = IPOIB_MIN_QUEUE_SIZE;
+ printk(KERN_WARNING "%s: ipoib_recvq_size is too small, use min %d instead\n", ca->name, IPOIB_MIN_QUEUE_SIZE);
+ }
+
+ priv->sendq_size = expsize(ipoib_sendq_size);
- priv->tx_ring = kzalloc(IPOIB_TX_RING_SIZE * sizeof (struct ipoib_tx_buf),
+ priv->tx_ring = kzalloc(priv->sendq_size * sizeof (struct ipoib_tx_buf),
GFP_KERNEL);
if (!priv->tx_ring) {
printk(KERN_WARNING "%s: failed to allocate TX ring (%d entries)\n",
- ca->name, IPOIB_TX_RING_SIZE);
+ ca->name, priv->sendq_size);
goto out_rx_ring_cleanup;
}
+ printk(KERN_INFO "%s: TX_RING_SIZE is set to %d entries\n",
+ ca->name, priv->sendq_size);
/* priv->tx_head & tx_tail are already 0 */
@@ -923,7 +968,7 @@
dev->hard_header_len = IPOIB_ENCAP_LEN + INFINIBAND_ALEN;
dev->addr_len = INFINIBAND_ALEN;
dev->type = ARPHRD_INFINIBAND;
- dev->tx_queue_len = IPOIB_TX_RING_SIZE * 2;
+ dev->tx_queue_len = priv->sendq_size * 2;
dev->features = NETIF_F_VLAN_CHALLENGED | NETIF_F_LLTX;
/* MTU will be reset when mcast join happens */
diff -urN infiniband/ulp/ipoib/ipoib_verbs.c infiniband-queue/ulp/ipoib/ipoib_verbs.c
--- infiniband/ulp/ipoib/ipoib_verbs.c 2006-03-26 11:57:15.000000000 -0800
+++ infiniband-queue/ulp/ipoib/ipoib_verbs.c 2006-03-31 08:46:34.308727224 -0800
@@ -159,8 +159,8 @@
struct ipoib_dev_priv *priv = netdev_priv(dev);
struct ib_qp_init_attr init_attr = {
.cap = {
- .max_send_wr = IPOIB_TX_RING_SIZE,
- .max_recv_wr = IPOIB_RX_RING_SIZE,
+ .max_send_wr = priv->sendq_size,
+ .max_recv_wr = priv->recvq_size,
.max_send_sge = 1,
.max_recv_sge = 1
},
@@ -175,7 +175,7 @@
}
priv->cq = ib_create_cq(priv->ca, ipoib_ib_completion, NULL, dev,
- IPOIB_TX_RING_SIZE + IPOIB_RX_RING_SIZE + 1);
+ priv->sendq_size + priv->recvq_size + 1);
if (IS_ERR(priv->cq)) {
printk(KERN_WARNING "%s: failed to create CQ\n", ca->name);
goto out_free_pd;
Thanks
Shirley Ma
IBM Linux Technology Center
15300 SW Koll Parkway
Beaverton, OR 97006-6063
Phone(Fax): (503) 578-7638
infiniband-tune-queue.patch
Description: Binary data
_______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
