x*********n
发帖数: 28013 | 1 发现still know nothing。
so sad | z**r
发帖数: 17771 | 2 都看的啥书啊？俺给客户讲4个小时QoS，俺看很多人都明白了不少
【在 x*********n 的大作中提到】
: 发现still know nothing。
: so sad
| x*********n
发帖数: 28013 | 3 就是看了你介绍的QOS for CCIP的那个wendell odom的书。
讲的太细了，tx-ring的那个limit啥的都提到了，
还有什么QPPB，这个居然是classification and marking，好难理解啊。
我觉得shape average和shape peak这个挺难理解的，打印了一个cisco learning的讨
论，准备慢慢看。
【在 z**r 的大作中提到】
: 都看的啥书啊？俺给客户讲4个小时QoS，俺看很多人都明白了不少
| z**r
发帖数: 17771 | 4 你可以先只看MQC的，其他的先别管。shaping 和 policing都要用到token bucket，这
个没办法，多看几遍理解吧
【在 x*********n 的大作中提到】
: 就是看了你介绍的QOS for CCIP的那个wendell odom的书。
: 讲的太细了，tx-ring的那个limit啥的都提到了，
: 还有什么QPPB，这个居然是classification and marking，好难理解啊。
: 我觉得shape average和shape peak这个挺难理解的，打印了一个cisco learning的讨
: 论，准备慢慢看。
| A***i
发帖数: 17 | 5 学简单点吧，Cisco好像有张QoS的cheat sheet.看着那个也就差不多了。
佩服Zher能讲4个小时。 | x*********n
发帖数: 28013 | 6 恩。多谢大哥。
single rate，dual rate，single bucket，dual bucket，
但是这还不是难点，难点是shape average 8000 8000 0，和shape peak x 有的时候算
出来是一样的，这个超难理解啊。以前读书的时候做过计算题也不知道当时是怎么做的。
【在 z**r 的大作中提到】
: 你可以先只看MQC的，其他的先别管。shaping 和 policing都要用到token bucket，这
: 个没办法，多看几遍理解吧
| x*********n
发帖数: 28013 | 7 QOS很烦的，特别是很多要配合到硬件，特别是到了high platform，全跟硬件搭配，一
下子推翻了很多理论。
不过大部分时候，估计MQC就more than enough了吧。
【在 A***i 的大作中提到】
: 学简单点吧，Cisco好像有张QoS的cheat sheet.看着那个也就差不多了。
: 佩服Zher能讲4个小时。
| p**x
发帖数: 123 | 8 讲4个小时挺多的 可以说说都什么topic吗？
【在 z**r 的大作中提到】
: 都看的啥书啊？俺给客户讲4个小时QoS，俺看很多人都明白了不少
|
| z**r
发帖数: 17771 | 9 IntServe
- RSVP
DiffServe
- Classification, IPP, COS, DSCP, IPv6 TC, MPLS
- Marking
-Explicit marking
-Implicit marking, policing set, trust domain, mpls modes, etc.
- Congestion management, PQ, CQ, LLQ, CBWFQ, scheduler, etc.
- Congestion avoidence, RED, WRED, FRED, TCP window
- Shaping and policing, token bucket, strict policer, conditional policer,
etc.
- MQC, H-QoS
- Implementation guide
【在 p**x 的大作中提到】
: 讲4个小时挺多的 可以说说都什么topic吗？
| p**x
发帖数: 123 | 10 lots of information in 4 hours ;)
I wonder what would be the relation in Diffserv and tcp windowing, you mean
using qos can actually improve the draw backs of tcp congestion avoidance
mechanism?
,
【在 z**r 的大作中提到】
: IntServe
: - RSVP
: DiffServe
: - Classification, IPP, COS, DSCP, IPv6 TC, MPLS
: - Marking
: -Explicit marking
: -Implicit marking, policing set, trust domain, mpls modes, etc.
: - Congestion management, PQ, CQ, LLQ, CBWFQ, scheduler, etc.
: - Congestion avoidence, RED, WRED, FRED, TCP window
: - Shaping and policing, token bucket, strict policer, conditional policer,
| | | z**r
发帖数: 17771 | 11 yes, in multiple tcp flows situation (this is true in the real networks),
randomly drop some tcp packets could improve the overall bandwidth
utilization
mean
【在 p**x 的大作中提到】
: lots of information in 4 hours ;)
: I wonder what would be the relation in Diffserv and tcp windowing, you mean
: using qos can actually improve the draw backs of tcp congestion avoidance
: mechanism?
:
: ,
| p**x
发帖数: 123 | 12 i agree randomly dropping packets can help in tcp environment but it is more
of a queuing strategy. this can cause problem as well, any of dropped
packet will need to reestablish transmission and go through slow start,
sliding window again...
【在 z**r 的大作中提到】
: yes, in multiple tcp flows situation (this is true in the real networks),
: randomly drop some tcp packets could improve the overall bandwidth
: utilization
:
: mean
| x*********n
发帖数: 28013 | 13 queuing is congestion management tool
shaping and policing is congestion avoidance tool
shaping is not just drop, it is actually buffer the packets.however,
policing is.
WRED has function such as mimimum threshold, and maximum threshold, between
them, there is probability to mark all the packets, based on dscp or ip
precedence, the chance of each packet to be drop is different,sounds like
smart, right?
I wish i could had chance to listen to zher's lecture, some topics I even
haven't heard of before, such as trust domain.
more
【在 p**x 的大作中提到】
: i agree randomly dropping packets can help in tcp environment but it is more
: of a queuing strategy. this can cause problem as well, any of dropped
: packet will need to reestablish transmission and go through slow start,
: sliding window again...
| x*********n
发帖数: 28013 | 14 什么是Explicit marking
-Implicit marking,啊？
,
【在 z**r 的大作中提到】
: IntServe
: - RSVP
: DiffServe
: - Classification, IPP, COS, DSCP, IPv6 TC, MPLS
: - Marking
: -Explicit marking
: -Implicit marking, policing set, trust domain, mpls modes, etc.
: - Congestion management, PQ, CQ, LLQ, CBWFQ, scheduler, etc.
: - Congestion avoidence, RED, WRED, FRED, TCP window
: - Shaping and policing, token bucket, strict policer, conditional policer,
| p**x
发帖数: 123 | 15 yeah too smart
between
【在 x*********n 的大作中提到】
: queuing is congestion management tool
: shaping and policing is congestion avoidance tool
: shaping is not just drop, it is actually buffer the packets.however,
: policing is.
: WRED has function such as mimimum threshold, and maximum threshold, between
: them, there is probability to mark all the packets, based on dscp or ip
: precedence, the chance of each packet to be drop is different,sounds like
: smart, right?
: I wish i could had chance to listen to zher's lecture, some topics I even
: haven't heard of before, such as trust domain.
| z**r
发帖数: 17771 | 16 slow start is the nature of tcp, without random dropping, it will still
happen. but with random drops, the overall bandwidth utilization is much
smoother, and more efficient. remember there are always many tcp sessions
concurrently
more
【在 p**x 的大作中提到】
: i agree randomly dropping packets can help in tcp environment but it is more
: of a queuing strategy. this can cause problem as well, any of dropped
: packet will need to reestablish transmission and go through slow start,
: sliding window again...
| p**x
发帖数: 123 | 17 "slow start is the nature of tcp, without random dropping, it will still
happen"
true. but these dropped packets should considered as a overhead for
retransmission and an extra process to establish communication again, if
they weren't dropped, they probably already made to the destination without
going through sliding window again...however, i've never done any poc on it,
it's just a thought that i'm hoping to get some truth behind it.
【在 z**r 的大作中提到】
: slow start is the nature of tcp, without random dropping, it will still
: happen. but with random drops, the overall bandwidth utilization is much
: smoother, and more efficient. remember there are always many tcp sessions
: concurrently
:
: more
| x*********n
发帖数: 28013 | 18 TCP is based on the client and server model,
How much you can send usually is based on receiver not sender, that's why
everywhere we need a notification.
To send efficiently is not the faster the better, is the fastest way to
reach the maximum rate to send,then it is the most efficient.
To find how fast start to reach that rate, we are using selective ACK.
Let me know if I am wrong, entry level like me are always confused by
similar term.
without
it,
【在 p**x 的大作中提到】
: "slow start is the nature of tcp, without random dropping, it will still
: happen"
: true. but these dropped packets should considered as a overhead for
: retransmission and an extra process to establish communication again, if
: they weren't dropped, they probably already made to the destination without
: going through sliding window again...however, i've never done any poc on it,
: it's just a thought that i'm hoping to get some truth behind it.
| z**r
发帖数: 17771 | 19 2 things you are missing:
1. there are always multiple tcp sessions in the real world
2. without congestion avoidance like RED, whenever there is a drop, it's
tail drop.
So, when the link is congested, all the packets that are not in the queue
yet will be tail dropped, and these packets are from multiple tcp sessions.
Thus results in transmission wave or global sync, big waste of bandwidth (
queue actually)
With congestion avoidance like RED/WRED/FRED/etc., the lower priority
packets with higher drop possibility (think about AF13 vs AF41 in DSCP
situation) will be dropped first once the minimum threshold is reached, so
the tcp sessions are differentiated and transmission wave is avoided.
Afterall, the overall bandwidth utilization would be much better.
without
it,
【在 p**x 的大作中提到】
: "slow start is the nature of tcp, without random dropping, it will still
: happen"
: true. but these dropped packets should considered as a overhead for
: retransmission and an extra process to establish communication again, if
: they weren't dropped, they probably already made to the destination without
: going through sliding window again...however, i've never done any poc on it,
: it's just a thought that i'm hoping to get some truth behind it.
| t*******r
发帖数: 3271 | | x*********n
发帖数: 28013 | 21 如果AF11 vs AF33，
哪个会被drop呢？
.
【在 z**r 的大作中提到】
: 2 things you are missing:
: 1. there are always multiple tcp sessions in the real world
: 2. without congestion avoidance like RED, whenever there is a drop, it's
: tail drop.
: So, when the link is congested, all the packets that are not in the queue
: yet will be tail dropped, and these packets are from multiple tcp sessions.
: Thus results in transmission wave or global sync, big waste of bandwidth (
: queue actually)
: With congestion avoidance like RED/WRED/FRED/etc., the lower priority
: packets with higher drop possibility (think about AF13 vs AF41 in DSCP
| p**x
发帖数: 123 | 22 Points well taken, thanks.
.
【在 z**r 的大作中提到】
: 2 things you are missing:
: 1. there are always multiple tcp sessions in the real world
: 2. without congestion avoidance like RED, whenever there is a drop, it's
: tail drop.
: So, when the link is congested, all the packets that are not in the queue
: yet will be tail dropped, and these packets are from multiple tcp sessions.
: Thus results in transmission wave or global sync, big waste of bandwidth (
: queue actually)
: With congestion avoidance like RED/WRED/FRED/etc., the lower priority
: packets with higher drop possibility (think about AF13 vs AF41 in DSCP
|
|
