Hi Sasha,
Sasha Khapyorsky wrote:
Hi Yevgeny,
Some initial comments.
On 01:07 Sun 22 Jul , Yevgeny Kliteynik wrote:
Hi All
Please find the attached RFC describing how QoS policy support could be
implemented in the OpenFabrics stack.
Your comments are welcome.
-- Yevgeny
RFC: OpenFabrics Enhancements for QoS Support
===============================================
Authors: . Eitan Zahavi <eitan at mellanox.co.il>
Authors: . Yevgeny Kliteynik <kliteyn at mellanox.co.il>
Date: .... Jul 2007.
Revision: 0.2
Table of contents:
1. Overview
2. Architecture
3. Supported Policy
4. CMA functionality
5. IPoIB functionality
6. SDP functionality
7. SRP functionality
8. iSER functionality
9. OpenSM functionality
1. Overview
------------
Quality of Service requirements stem from the realization of I/O
consolidation
over IB network: As multiple applications and ULPs share the same fabric,
means
to control their use of the network resources are becoming a must. The basic
need is to differentiate the service levels provided to different traffic
flows,
such that a policy could be enforced and control each flow utilization of
the
fabric resources.
IBTA specification defined several hardware features and management
interfaces
to support QoS:
* Up to 15 Virtual Lanes (VL) carry traffic in a non-blocking manner
* Arbitration between traffic of different VLs is performed by a 2 priority
levels weighted round robin arbiter. The arbiter is programmable with
a sequence of (VL, weight) pairs and maximal number of high priority
credits
to be processed before low priority is served
* Packets carry class of service marking in the range 0 to 15 in their
header SL field
* Each switch can map the incoming packet by its SL to a particular output
VL based on programmable table VL=SL-to-VL-MAP(in-port, out-port, SL)
* The Subnet Administrator controls each communication flow parameters
by providing them as a response to Path Record (PR) or MultiPathRecord
(MPR)
queries
The IB QoS features provide the means to implement a DiffServ like
architecture.
DiffServ architecture (IETF RFC2474 2475) is widely used today in highly
dynamic
fabrics.
This proposal provides the detailed functional definition for the various
software elements that are required to enable a DiffServ like architecture
over
the OpenFabrics software stack.
2. Architecture
----------------
This proposal split the QoS functionality between the SM/SA, CMA and the
various
ULPS. We take the "chronology approach" to describe how the overall system
works:
2.1. The network manager (human) provides a set of rules (policy) that
defines
how the network is being configured and how its resources are split to
different
QoS-Levels. The policy also define how to decide which QoS-Level each
application or ULP or service use.
2.2. The SM analyzes the provided policy to see if it is realizable and
performs
the necessary fabric setup. The SM may continuously monitor the policy and
adapt
to changes in it. Part of this policy defines the default QoS-Level of each
partition. The SA is being enhanced to match the requested Source,
Destination,
QoS-Class, Service-ID (and optionally SL and priority) against the policy.
So
clients (ULPs, programs) can obtain a policy enforced QoS. The SM is also
enhanced to support setting up partitions with appropriate IPoIB broadcast
group. This broadcast group carries its QoS attributes: SL, MTU and
RATE.
2.3. IPoIB is being setup. IPoIB uses the SL, MTU and RATE available on the
multicast group which forms the broadcast group of this partition.
2.4. MPI which provides non IB based connection management should be
configured
to run using hard coded SLs. It uses these SLs for every QP being opened.
2.5. ULPs that use CM interface (like SRP) should have their own
pre-assigned
Service-ID and use it while obtaining PR/MPR for establishing connections.
The SA receiving the PR/MPR should match it against the policy and return
the appropriate PR/MPR including SL, MTU and RATE.
2.6. ULPs and programs using CMA to establish RC connection should provide
the
CMA the target IP and Service-ID. Some of the ULPs might also provide
QoS-Class
(E.g. for SDP sockets that are provided the TOS socket option). The CMA
should
then use the provided Service-ID and optional QoS-Class and pass them in the
PR/MPR request. The resulting PR/MPR should be used for configuring the
connection QP.
PathRecord and MultiPathRecord enhancement for QoS:
As mentioned above the PathRecord and MultiPathRecord attributes should be
enhanced to carry the Service-ID which is a 64bit value, which has been
standardized by the IBTA. A new field QoS-Class is also provided.
A new capability bit should describe the SM QoS support in the SA class port
info. This approach provides an easy migration path for existing access
layer
and ULPs by not introducing new set of PR/MPR attribute.
3. Supported Policy
--------------------
The QoS policy supported by this proposal is divided into 4 sub sections:
I) Port Group: a set of CAs, Routers or Switches that share the same
settings.
A port group might be a partition defined by the partition manager policy in
terms of GUIDs. Future implementations might provide support for
NodeDescription
based definition of port groups.
Isn't it better to have port group definitions in separate file? So
groups could be shared with other OpenSM components (as discussed). Even
if such group sharing is not high priority functionality this should
save us from redoing things later.
II) Fabric Setup:
Defines how the SL2VL and VLArb tables should be setup. This policy
definition
assumes the computation of overall end to end network behavior should be
performed
outside of OpenSM.
III) QoS-Levels Definition:
This section defines the possible sets of parameters for QoS that a client
might be mapped to. Each set holds: SL and optionally: Max MTU, Max Rate,
Packet Lifetime and Path Bits (in case LMC > 0 is used for QoS).
IV) Matching Rules:
A list of rules that match an incoming PR/MPR request to a QoS-Level. The
rules are processed in order such as the first match is applied. Each rule
is
built out of a set of match expressions which should all match for the rule
to
apply. The matching expressions are defined for the following fields
** SRC and DST to lists of port groups
** Service-ID to a list of Service-ID or Service-ID ranges
** QoS-Class to a list of QoS-Class values or ranges
QoS Policy file syntax
* Empty lines are ignored
* Leading and trailing blanks, as well as empty lines, are ignored, so the
indentation in the example is just for better readability
* Comments are started with the pound sign (#) and terminated by EOL
* Comments may appear only in a separate line
Why? What is wrong with:
port-name: vs1/HCA-1/P1 # my best port
I can use this too, but then the pound sign, wherever it will
appear, would mean commentary start. No \# or something like this
to include it in some other place - I don't want to complicate the
syntax. Sounds OK?
* Keywords that denote section/subsection start have matching closing
keywords
* Any keyword should be the first non-blank in the line
QoS Policy file example
# Port Groups define sets of ports to be used later in the settings
port-groups
# using port GUIDs
port-group
name: Storage
# "use" is just a description that is used for logging.
# Other than that, it is just a commentary
use: our SRP storage targets
port-guid: 0x1000000000000001
port-guid: 0x1000000000000002
end-port-group
port-group
name: Virtual Servers
use: node desc and IB port num
# The syntax of the port name is as follows:
"hostname/CA-num/Pnum".
# "hostname" and "CA-num" are compared to the first 2 words of
# NodeDescription, and "Pnum" is a port number on that node.
port-name: vs1/HCA-1/P1
port-name: vs3/HCA-1/P1
port-name: vs3/HCA-2/P2
What about wild carding here, like vs1/*/* or just vs1?
Good idea.
end-port-group
# using partitions defined in the partition policy
port-group
name: Group for Partition 1
use: default settings
partition: Part1
end-port-group
# using node types CA|ROUTER|SWITCH
Probably also ALL (for all ports), SELF (for SM port)?
Agree.
port-group
name: Routers
use: all routers
node-type: ROUTER
end-port-group
end-port-groups
I agree that proposed syntax has better for human readability than pure
XML, but isn't stuff like this will be more user-friendly?
Storage "Free Text description" = 0x10001, 0x10002, 0x10003 ;
, or
Storage "Free Text description" { 0x10001, 0x10002, 0x10003 };
, or
Storage "Free Text description": ROUTERS, CAS ;
GUID list is a good idea.
Not sure about the other stuff. A certain port group can be defined
both by guids and by node-types. How about this:
port-group
name: routers_and_mgt_nodes
use: all routers and management nodes
node-type: ROUTER
port-guid: 0x10001, 0x10002, 0x10003
end-port-group
qos-setup
# define all types of VLArb tables. The length of the tables should
# match the physically supported tables by their target ports
vlarb-tables
# scope defines the exact ports the VLArb tables apply to
vlarb-scope
# defining VLArb tables on all the ports that belong to
# port group 'Storage', and on all the ports connected
# to ports of port group 'Storage'
group: Storage
So "group" is only for ports that belong to 'Storage'?
Yes, and "across" is for ports that connected to ports of group 'Storage'
# "across" means all the ports that are connected to ports
# that belong to the specified port group
across: Storage
# VLArb table holds VL and weight pairs
vlarb-high: 0:255,1:127,2:63,3:31,4:15,5:7,6:3,7:1
vlarb-low: 8:255,9:127,10:63,11:31,12:15,13:7,14:3
vl-high-limit: 10
end-vlarb-scope
# There can be several scopes
end-vlarb-tables
sl2vl-tables
# Scope defines the exact devices and in/out ports tables apply
to.
# Note: if the same port is matching several rules the *FIRST*
one applies.
sl2vl-scope
# SL2VL tables are orgnized as SL2VL(in-port,out-port)
# "from: n,m" means we define the SL2VL(n,*) and SL2VL(m,*)
# "to: n,m" means we define the SL2VL(*,n) and SL2VL(*,m)
#
# The following example specifies that all the SL2VL tables
# entries should be defined for all the ports of group
Part1:
group: Part1
from: *
to: *
# SL2VL table has to have 16 values at max - one for each
SL.
# If the user specifies less than 16 values, all the missing
# VL values will be implicitly set to 0
sl2vl-table: 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,7
end-sl2vl-scope
sl2vl-scope
# "across-to" is a combination of "across" keyword
(definition can be found
# in VLArb tables section) and "to" keyword.
# "across: PortGroupName" refers to all the ports that are
connected
# to ports that belong to PortGroupName.
#
# Example of "across-to" usage:
# A user has a set of 'special' nodes (e.g. storage
nodes), and all
# the traffic to these nodes has to get specific VL.
# The solution is to define port group (i.g. "Storage")
that will
# include all the ports of these nodes, and then to
configure SL2VL
# tables on all the switch ports that are connected to the
Storage
# port group by specifying "across-to: Storage".
#
across-to: Storage2
# Similar to "across-to", "across-from" is a combination of
"across"
# and "to" keywords
across-from: Storage1
sl2vl-table: 0,1,1,1,1,1,1,1,1,1,1,1,1,1,1,0
end-sl2vl-scope
end-sl2vl-tables
end-qos-setup
qos-levels
# the first one is just setting SL
qos-level
use: for the lowest priority communication
sl: 15
packet-life: 16
end-qos-level
# the second sets SL and QoS Class
qos-level
use: low latency best bandwidth
sl: 0
end-qos-level
# the whole set: SL, MTU-Limit, Rate-Limit, Packet Lifetime, Path
Bits
qos-level
use: just an example
sl: 0
mtu-limit: 1
rate-limit: 1
packet-life: 12
# Path Bits can be used e.g. to provide a different routes
through the
# subnet to a particular port
path-bits: 2,4,8-32
end-qos-level
end-qos-levels
# Match rules are scanned in a first-fit manner (like firewall rules
table)
qos-match-rules
# matching by single criteria: class (list of values and ranges)
qos-match-rule
# just a description
use: low latency by class 7-9 or 11
qos-class: 7-9,11
# number of qos-level to apply to the matching PR/MPR
qos-level-sn: 1
Isn't it better and less error prone to match qos_level by name and not
by sequential number?
qos-level can have name, and then qos-match-rule will refer to this name.
But matching qos-level by sequential number makes it really easy to locate
the referred qos-level, which is important, as every PR/MPR request would
go through this process, so saving some runtime in this area is important IMHO.
end-qos-match-rule
# show matching by destination group AND service-ids
qos-match-rule
use: Storage targets connection
destination: Storage
service-id: 22,4719-5000
qos-level-sn: 2
end-qos-match-rule
# show matching by source group only
qos-match-rule
use: bla bla
source: Storage
qos-level-sn: 3
end-qos-match-rule
end-qos-match-rules
4. IPoIB
---------
IPoIB already query the SA for its broadcast group information. The
additional
functionality required is for IPoIB to provide the broadcast group SL, MTU,
and RATE in every following PathRecord query performed when a new UDAV is
needed by IPoIB.
We could assign a special Service-ID for IPoIB use but since all
communication
on the same IPoIB interface shares the same QoS-Level without the ability to
differentiate it by target service we can ignore it for simplicity.
5. CMA features
----------------
The CMA interface supports Service-ID through the notion of port space as a
prefixes to the port_num which is part of the sockaddr provided to
rdma_resolve_add(). What is missing is the explicit request for a QoS-Class
that
should allow the ULP (like SDP) to propagate a specific request for a class
of
service. A mechanism for providing the QoS-Class is available in the IPv6
address,
so we could use that address field. Another option is to implement a special
connection options API for CMA.
Missing functionality by CMA is the usage of the provided QoS-Class and
Service-ID
in the sent PR/MPR. When a response is obtained it is an existing
requirement for
the CMA to use the PR/MPR from the response in setting up the QP address
vector.
6. SDP
-------
SDP uses CMA for building its connections.
The Service-ID for SDP is 0x000000000001PPPP, where PPPP are 4 hex digits
holding the remote TCP/IP Port Number to connect to.
SDP might be provided with SO_PRIORITY socket option. In that case the value
provided should be sent to the CMA as the TClass option of that connection.
7. SRP
-------
Current SRP implementation uses its own CM callbacks (not CMA). So SRP
should
fill in the Service-ID in the PR/MPR by itself and use that information in
setting up the QP. The T10 SRP standard defines the SRP Service-ID to be
defined
by the SRP target I/O Controller (but they should also comply with IBTA
Service-
ID rules). Anyway, the Service-ID is reported by the I/O Controller in the
ServiceEntries DMA attribute and should be used in the PR/MPR if the SA
reports its ability to handle QoS PR/MPRs.
8. iSER
--------
iSER uses CMA and thus should be very close to SDP. The Service-ID for iSER
should be TBD.
9. OpenSM features
-------------------
The QoS related functionality to be provided by OpenSM can be split into two
main parts:
3.1. Fabric Setup
During fabric initialization the SM should parse the policy and apply its
settings to the discovered fabric elements. The following actions should be
performed:
* Parsing of policy
* Node Group identification. Warning should be provided for each node not
specified but found.
* SL2VL settings validation should be checked:
+ A warning will be provided if there are no matching targets for the
SL2VL
setting statement.
+ An error message will be printed to the log file if an invalid setting
is
found. A setting is invalid if it refers to:
- Non existing port numbers of the target devices
- Unsupported VLs for the target device. In the later case the map to
non
existing VLs should be replaced to VL15 i.e. packets will be dropped.
I'm not sure it is optimal. We could have well documented or even
configurable mapping rule instead, then this will not limit devices with
higher capabilities.
I'm open for suggestions.
* SL2VL setting is to be performed
* VL Arbitration table settings should be validated according to the
following
rules:
+ A warning will be provided if there are no matching targets for the
setting
statement
+ An error will be provided if the port number exceeds the target ports
+ An error will be generated if the table length exceeds device
capabilities
Ditto.
+ A warning will be generated if the table quote a VL that is not supported
by the target device
What is "table quote" here?
* VL Arbitration tables will be set on the appropriate targets
3.2. PR/MPR query handling:
OpenSM should be able to enforce the provided policy on client request.
The overall flow for such requests is: first the request is matched against
the
defined match rules such that the target QoS-Level definition is found.
Given
the QoS-Level a path(s) search is performed with the given restrictions
imposed
by that level. The following two sections describe these steps.
How Service-ID is carried in the PathRecord and MultiPathRecord attributes
is
now standardized by the IBTA.
3.2.1. Matching rule search:
A rule is "matching" a PR/MPR request using the following criteria:
* Matching rules provide values in a list of either single value, or range
of
values. A PR/MPR field is "matching" the rule field if it is explicitly
noted in the list of values or is one of the values covered by a range
included in the field values list.
* Only PR/MPR fields that have their component mask bit set should be
compared.
* For a rule to be "matching" a PR/MPR request all the rule fields should be
"matching" their PR/MPR fields. Such that a PR/MPR request that does
not have a component mask field set for one of the rule defined fields
can
not match that rule.
* A PR/MPR request that have a component mask bit set for one of the fields
that is not defined by the rule can match the rule.
Aren't last two too restrictive? SA can just to filter-out paths in
response to match rest of the rule. No?
Not sure I'm following.
The last bullet is not restrictive at all - it says that if you have a match
rule with some reduced set of fields (e.g. only service id), any PR/MPR with
a matching service id will be matched, even if it also has MTU, rate, etc.
The algorithm to be used for searching for a rule match might be as simple
as a
sequential search through all rules or enhanced for better performance. The
semantics of every rule field and its matching PR/MPR field are described
below:
* Source: the SGID or SLID should be part of this group
* Destination: the DGID or DLID should be part of this group
* Service-ID: check if the requested Service-ID (available in the PR/MPR old
SM-Key field) is matching any of this rule Service-IDs
* TClass: check if the PR/MPR TClass field is matching
3.2.2 PR/MPR response generation:
The QoS-Level pointed by the first rule that matches the PR/MPR request
should be used for obtaining the response SL, MTU-Limit, RATE-Limit,
Path-Bits
and QoS-Class. A default QoS-Level should be used if no rule is matching the
query.
Where this default should be defined?
OK, I missed that part. Here it is:
- qos-level sequential number is counted from 0
- qos-level num. 0 is a must is treated as the Default Level - it's
applied to any PR/MPR request that didn't match any match rule
- default qos-level can be also referred explicitly in any match rule
by specifying "qos-level-sn: 0"
-- Yevgeny
Sasha
The efficient algorithm for finding paths that meet the QoS-Level criteria
is
beyond the scope of this RFC and left for the implementer to provide.
However
the criteria by which the paths match the QoS-Level are described below:
* SL: The paths found should all use the given SL. For that sake PR/MPR
algorithm should traverse the path from source to destination only through
ports that carry a valid VL (not VL15) by the SL2VL map (should consider
input
and output ports and SL).
* MTU-Limit: The resulting paths MTU should not exceed the given MTU-Limit
* Rate-Limit: The resulting paths RATE should not exceed the given
RATE-Limit
(rate limit is given in units of link BW = Width*Speed according to IBTA
Specification Vol-1 table-205 p-901 l-24).
* Path-Bits: define the target LID lowest bits (number of bits defined by
the
target port PortInfo.LMC field). The path should traverse the LFT using
the
target port LID with the path-bits set.
* QoS-Class: should be returned in the result PR/MPR. When routing is going
to
be supported by OpenSM we might use this field in selecting the target
router too in a TBD way.
_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit
http://openib.org/mailman/listinfo/openib-general
_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
