You are on page 1of 80

1Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

SPARC T5 Servers Deep Dive


NDA Part 1

Drag picture to placeholder


or click icon to add

Insert Presenters Name Here


Insert Presenters Title Here

Drag picture to placeholder or click icon

2Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

The following is intended to outline our general product direction. It is


intended for information purposes only,
and may not be incorporated into any contract. It is
not a commitment to deliver any material, code, or functionality, and
should not be relied upon in making purchasing decisions.
The development, release, and timing of any features or functionality
described for Oracles products remains at the sole discretion of Oracle.

3Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Design Objectives Achieved


Oracle workloads

Engineered Systems

Extends

on-chip crypto
acceleration
RAS

Scales to 8 sockets
using directory

Minimizes latency

Avoids congestion

Maximize bandwidth

4Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Double cores and


cache

Balance single thread


and throughput

Dynamically thread

Optimize
Systems

Multiply
Performance
SPARC
T5

Scale
Efficiently

Confidential Oracle Internal

Advance
Power
Management

Maximizes peak
performance

Manages thermal and


current loads

Scales elastically

Next Generation SPARC T5 Servers


Faster. Optimized. Secure.

SPARC T5-1B

SPARC T5-4

SPARC T5-2

5Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

SPARC T5-8

SPARC T5 Servers
Product Line Overview

Processor

SPARC T5-1B

SPARC T5-2

SPARC T5-4

SPARC T5-8

SPARC T5 3.6GHz

SPARC T5
3.6GHz

SPARC T5
3.6GHz

SPARC T5
3.6GHz

16, 128

32, 256

64, 512

128, 1024

16

32

64

128

128GB or 256GB

256GB or 512GB

1TB or 2TB

2TB or 4TB

16 LP x8 PCIe 3.0,
4 x 10GbE ports

16 LP x8 PCIe 3.0,
4 x 10GbE ports

Rack 5RU
2410W

Rack 8RU
4850W

Max Processor Chips


Max Cores/Threads
DIMM Slots
Max Memory
Drive Bays
I/O Slots
Form Factor/RU
Max Power Consumption

2 x PCIe 2.0 EM,


2 NEM,1 FEM slots

8 LP x8 PCIe 3.0,
4 x 10GbE ports

Blade
689W

6Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Rack 3RU
1927W

New l SPARC T5-1B Blade Server


Next Generation 1-socket SPARC Blade Server

Compute
1x SPARC T5 16-core CPU
16x DDR3 DIMMs
128GB (8GB DIMMs) or 256GB (16GB DIMMs)
memory

I/O and Storage


2x hot-plug PCIe 2.0 x8 Express Modules
2x 2.5 SAS HDD or SSD drives

Availability and Management


Built-in RAID 0, 1, 1E
Hot plug disks
Integrates with Sun Blade 6000 network architecture
Oracle ILOM service processor

7Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5-1B Board
BoBs
REM
T5 CPU

PCIe
Switch

Service
Processor

16 DDR3
DIMMs
8Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

New l SPARC T5-2 Enterprise Server


Next Generation 2-socket SPARC Server

Compute
2x SPARC T5 16-core CPU
32x DDR3 DIMMs
256GB (8GB DIMMs) or 512GB (16GB DIMMs) of
memory

I/O and Storage


8x PCIe 3.0 x8 slots
4x 10G-baseT ports
6x 2.5 SAS HDD or SSD drives

Availability and Management


Built-in RAID 0, 1, 1E
Hot-plug disks
Hot-swap and redundant fans and power supplies
Oracle ILOM service processor

9Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5-2 Chassis
T5 CPU

Memory
Risers

Service
Processor

10Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5-2 Front
Locator
LED/Button

Power
Button

Fault LED
Status
LED

Over Temp
Indicator
Status LEDs

DISK 5
DISK 4
DISK 3
DISK 2
DISK 1
DISK 0

RFID/Serial
Number

HD-15 VGA
Port
2x USB 3.0
Ports

11Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

DVD

T5-2 Rear
HD-15 VGA
Port

SP Network 10/100

12Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

4x 10GbE
Ports

2x USB 3.0

PCIe 8

System Rear
Indicators

PCIe 7

AC0

PCIe 6

PCIe 5

PCIe 4

PCIe 3

PCIe 2

PCIe 1

PSU 0

PSU 1
AC1

SP Serial

New l SPARC T5-4 Enterprise Server


Next Generation 4-socket SPARC Server
Compute
4x SPARC T5 16-core CPU
64x DDR3 DIMMs
1TB (16GB DIMMs) or 2TB (32GB DIMMs) memory

I/O and Storage


16x PCIe 3.0 x8 slots, w/ carriers
4x 10G-baseT ports
8x 2.5 SAS HDD or SSD drives

Front

Availability and Management


Built-in RAID 0, 1, 1E
Hot-plug disks, PCI cards
Hot-swap and redundant fans and power supplies
Oracle ILOM service processor

Rear
13Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

SPARC T4-4 Comparison SPARC to T5-4

Feature
Form Factor
CPU
Memory

SPARC T5-4
5RU, 31.5 deep
4x SPARC T5
3.6 GHz (512 threads)
DDR3, 2TB MAX
64x Slots

Network
Internal Storage
Removable Media
Serial
PCI Express slots
Power Supply
Fans

4x 10GbE
Up to 8 x 2.5 SAS 3.0 or SSD, hot-plug

SPARC T4-4
5RU, 28 deep
4x SPARC T4
3.0 GHz (256 threads)
DDR3, 2TB MAX
64x slots
4 x 1GbE + 8x 10GbE (XAUI)
Requires 2 Separate QSFP Connectors
Up to 8 x 2.5 SAS 2.0, can use up to
4x SATA SSDs, hot-plug

1x DVD-RW (via rKVMS; not local)

1x DVD-RW (via rKVMS; not local)

1x RS-232, 4x USB, 1x VGA


16x x8 slots (Hot-Plug Low Profile slots
with carrier card, ver 3.0)
2 x 3000 Watt AC, N+N
Redundant/Hot-Swap
5 x Redundant Hot-Swap

1x RS-232, 4x USB, 1x VGA


16x x8 slots (Hot-Plug Express Module slots,
ver 2.0)
4 x 2060 Watt AC, N+N
Redundant/Hot-Swap
5 x Redundant Hot-Swap

14Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

SPARC T5-4 Front Panel

Processor
Module

Processor Module Status LEDs


Locator
LED/Button

PM 1
PM 0

Fault LED
Status LED
Power
Button

DISK 1
DISK 0

Over Temp
Indicator

Rear Fan/EM
Indicator

RFID/Serial
Number

DISK 3
DISK 2

DISK 5
DISK 4

PSU 0
SP Serial
Port

15Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Main
Module

DISK 7
DISK 6

PSU 1
HD-15 VGA
Port

Confidential Oracle Internal

Dual USB 2.0


Ports

Main Module &


SP Status LEDs

Main Module
(Entire Board)

PSU Status
LEDs

SPARC T5-4 Rear Panel


Fan Module 0

Fan Module 2

Fan Module 3

Fan Module 4

PCIe 16

PCIe 15

PCIe 14

PCIe 13

PCIe 12

PCIe 11

PCIe 10

PCIe 9

PCIe 8

PCIe 7

PCIe 6

PCIe 5

PCIe 4

PCIe 3

PCIe 2

PCIe 1

Rear I/O
Module
(RIO)

Fan Module 1

PCIe Carrier
Hot-Plug Button, LEDs
AC3 OK
LED

AC0 OK LED

AC3
System Rear
Indicators

C19 Plug

AC0
SP Serial

16Copyright 2012, Oracle and/or its affiliates. All rights reserved.

4x 10GbE Ports
SP Network 10/100

Confidential Oracle Internal

2x USB 3.0

HD-15 VGA Port

C19 Plug

New l SPARC T5-8 Enterprise Server


Next Generation 8-socket SPARC Server

Compute
8x SPARC T5 16-core CPU
128x DDR3 DIMMs
2TB (16GB DIMMs) or 4TB (32GB DIMMs) memory

I/O and Storage


16x PCIe 3.0 x8 slots, w/ carriers
4x 10G-baseT ports
8x 2.5 SAS HDD or SSD drives

Front

Availability and Management


Built-in RAID 0, 1, 1E
Hot-plug disks, PCI cards
Hot-swap and redundant fans and power supplies
Oracle ILOM service processor

Rear
17Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Comparing T3, T4, and T5


Processor Clock
# of cores
Core Architecture
# of memory controllers

DIMMs/BoB
PCIe Gen support

PCIe Card Form Factor

18Copyright 2012, Oracle and/or its affiliates. All rights reserved.

SPARC T3

SPARC T4

SPARC T5

1.65GHz

2.85GHz, 3.0GHz

3.6GHz

16

16

S2

S3
2

2 BoBs/memory controller
4 DIMMs/BoB
4 BoBs/CPU socket
2.0

2 BoBs/memory controller
2 DIMMs/BoB
8 BoBs/CPU socket
3.0

Blade: Express Modules


1 and 2 socket rack server: LP
4 socket server: Express Modules

Blade: Express Modules


1-2 socket rack server: LP
4-8 socket rack server: LP on carrier card

Confidential Oracle Internal

T5-4 and T5-8 Processor Module


2 BoBs per Memory
Controller in the T5

T5 Processor 0

16 DIMMs per T5
16 DIMMs per T5

T5 Processor 1

19Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5-8 Processor Connectivity


T5-8 (8P option)

T5-8 (6P option)

1 CL between the 8 nodes

CL routing
6

3
5

3
5

0 1

C1

PM0
C2

C3

C4

C5

C6

0 1

C2

C0

Part #
Sw 0

Partition #
Switch 1

Partition #
Switch 2

C6
5

C1

2
3

C3

C4

C1

C5

C6

PM3
C6 C7
0 1 0 1

PMO
C0 C1
0 1 0 1

PM1
C2 C3
0 1 0 1

C0

C7

C0

Partition #
Switch 2

Part #
Sw 4

Part #
Sw 0

PCIe switches on MB in common


T5-4/8 Main Module

20Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Partition #
Switch 1

Partition #
Switch 2

PFM (not required)


Or Empty PM slot

PFM (not required)


Or Empty PM slot

PFM1

PFM1

C1

C2

C3

C1

PM0
C4

C5

C6

C0

C7

C0

PMO
C0 C1
0 1 0 1

PFM1

PFM2

C1

PM0
C2

C3

2
Partition #
Switch 2

PCIe switches on MB in common


T5-4/8 Main Module

Confidential Oracle Internal

1
Part #
Sw 4

0
Part #
Sw 0

3
Partition #
Switch 1

C5

C6

C7

PMO
C0 C1
0 1 0 1

PFM1

PFM2

PFM3

MP8?
3

Partition #
Switch 2

C1

PCIe upstream routing (PSR)


PM3
C6 C7
0 1 0 1

MP8?
3

C4

SSI, Present, 12C


(FPGA connectivity)
FPGA-T5-48

PCIe upstream routing (PSR)


PM3
C6 C7
0 1 0 1

0
4

SSI, Present, 12C


(FPGA connectivity)
FPGA-T5-48

PFM2

PFM2

MP8?
3

C7

4 5

4
2

PM0
C2

C1

PCIe upstream routing (PSR)

PM2
C4 C5
0 1 0 1

5
2

MP8?
0

SSI, Present, 12C


(FPGA connectivity)
FPGA-T5-48

PCIe upstream routing (PSR)


PM1
C2 C3
0 1 0 1

C0

C7

PM1

4
2

SSI, Present, 12C


(FPGA connectivity)
FPGA-T5-48

PMO
C0 C1
0 1 0 1

C7

4 5

PFM2

PFM not connected (required for airflow/EMI only

C1

2
3

2 CL between the 2 nodes

CL routing

PM3

C6
5

C1

C0

5
2

C5

C0

6
2
5 4

PM1

C7

4 5

PM2

C4

C2

2
3

T5-8 (2P option)

2 CL between the 4 nodes

CL routing

PM3

C6
5

T5-8 (4P option)

1 CL between the 6 nodes

CL routing

PM3

Note: the T5-8 2P option is


not a shipping configuration

3
Partition #
Switch 2

PCIe switches on MB in common


T5-4/8 Main Module

1
Part #
Sw 4

0
Part #
Sw 0

0
Partition #
Switch 1

0
Partition #
Switch 2

Partition #
Switch 2

PCIe switches on MB in common


T5-4/8 Main Module

Part #
Sw 4

T5-4 Processor Connectivity


T5-8 (4P option)

T5-8 (2P option)

2 CL between the 4 nodes

CL routing

2 CL between the 2 nodes

CL routing

PM1
6

C6
5

C0

C0

2
3

C1

C2

C3

C4

C1

PM0

PFM1

C7

4 5

C5

C6

C0

C7

C0

C1

SSI, Present, 12C


(FPGA connectivity)
FPGA-T5-48

PM0

C2

C3

Part #
Sw 0

3
Partition #
Switch 1

3
Partition #
Switch 2

PCIe switches on MB in common


T5-4/8 Main Module

21Copyright 2012, Oracle and/or its affiliates. All rights reserved.

C5

C6

C7

PMO
C0 C1
0 1 0 1

PFM1

MP4
3

Partition #
Switch 2

C1

PCIe upstream routing (PSR)


PM3
C6 C7
0 1 0 1

MP4
0

C4

SSI, Present, 12C


(FPGA connectivity)
FPGA-T5-48

PCIe upstream routing (PSR)


PMO
C0 C1
0 1 0 1

0
4

Confidential Oracle Internal

1
Part #
Sw 4

0
Part #
Sw 0

0
Partition #
Switch 1

0
Partition #
Switch 2

Partition #
Switch 2

PCIe switches on MB in common


T5-4/8 Main Module

Part #
Sw 4

T5-1B Block Diagram


DIMM DIMM DIMM DIMM DIMM DIMM DIMM DIMM

BoB

BoB

BoB

CPU
DC/DCs

BoB

BoB

T5
CPU0

BoB

BoB

BoB

Host &
Data Flash

TPM

CPU
Debug
Port

PCIe
0
1

FPGA

Disk1
Disk2

x2
x1

LSI SAS

x8

x8

PCIe Gen3
PCIe Gen2

x8
x1
PCIe
Switch 0

DBG

PCIe
Switch 1

LPC

SP Module

Sideband Mgmt

x4

x8

x8

FEM0

USB 3.0 USB


USB 2.0
1x4

PCI-EM1

1x4

1x4

Confidential Oracle Internal

USB 2.0 Enet


Storage

Emulex
Pilot 3

USB 1.0
Hub Ctrl

USB

VGA
Serial
(HD15) (RJ45)

Front UCP (Dongle Cable)

1x4

NEM0 NEM0 NEM1 NEM1


(4:7) (0:3) (0:3) (4:7)

? here
22Copyright 2012, Oracle and/or its affiliates. All rights reserved.

USB 3.0
Host Ctrl

FEM1
Nalia
Niantic

Dual GigE
10/100/1000

NEM0 NEM1

USB 1.1 Keyboard/Mouse

x8

12C

PCI-EM0

MIDPLANE

Ethernet Mgmt
(to CMM)

T5-2 Block Diagram


DIMM

DIMM

DIMM

DIMM

DIMM

DIMM

DIMM

DIMM

DIMM

DIMM

DIMM

DIMM

DIMM

DIMM

DIMM

DIMM

BoB

BoB

BoB

BoB

BoB

BoB

BoB

BoB

BoB

BoB

BoB

BoB

BoB

BoB

BoB

BoB

T5-0

T5-1
CPU
Debug
Port

PCIe

CPU
DC/DCs

CPU
DC/DCs

TPM

x8

x4

x8
x1

HDD0

SAS/SATA
IO Controller

HDD0
HDD0

x1
x4

PCIe
Switch 0

HDD0
HDD0

SP Module

PCIe
Switch 1

x8

FRUID
Sideband Mgmt

x8

x8

HDD0

USB 3.0
Host

SATA DVD

USB 3.0
Hub

DBG

x4
x8

get rid of all inside


small boxes

SAS/SATA
IO Controller

Host &
Data Flash

FPGA

x8

x8

CPU
Debug
Port

PCIe

USB 2.0 Storage

DRAM
SPI
Flash
NAND

USB 2.0
Hub

USB

USB
VGA

Service
Processor

USB 1.1 Keyboard Mouse

Internal USB
VGA

FAN BOARD

REAR IO

23Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Slot 8 (8)

Slot 7 (8)

Slot 6 (8)

Quad 10Gig Enet

Slot 5 (8)

Slot 4 (8)

Slot 3 (8)

Slot 2 (8)

USB3 VGA

Slot 1 (8)

REAR IO Board
USB2

USB0 USB1

VGA
DB15

Serial
Mgmt

Enet
Mgmt
10/100

T5-4 & T5-8 I/O Block Diagram


HDD[4:7]

SBP1

4-DISK BACKPLANE
DISKS 4-7

MB
Motherboard

2x USB (Gen3)
[VGA/DB15]
[Serial Mgmt]

HDD[0:3]
FRONT IO
VGA, SERIAL MGMT, USB

TO RIO
SAS 2308
SAS1

4-DISK BACKPLANE
DISKS 0-3

FIO + VGA

VIDEO
MUX

Serial
MUX

SP

SERVICE PROCESSOR

MM
Main Module

SBP0

TO RIO

x2 USB Ports
SAS 2308
SAS0

Enet Mgmt
NC-SI

FRONT USB
HOST CTRLLR.

FPGA, TOD, TPM

MONITOR &
CONTROL

SW1

CPU
PCIE PORTS

1 2

REAR USB
MOST CTRLLR.

10GB NIC 0

1x USB (Gen2)

2x USB (Gen3)

24Copyright 2012, Oracle and/or its affiliates. All rights reserved.

4 5 6

SW3

SW2

CPU
PCIE PORTS

SW5

CPU
PCIE PORTS

7 8 9 10

RIO Rear IO Module


Quad 10Gig Enet (10GBase-T Copper)

Confidential Oracle Internal

CPU
PCIE PORTS

11 12 13 14

SW4

Net 1

Net 0

RIO

SW0

CLOCK SYNTH.
&
BUFFERS

Debug Conn

Debug Conn

DC-DC
CONVERTERS

SW6

CPU
PCIE PORTS

EB
PCI Express Backplane

10GB NIC 1

15 16

[VGA/DB15]
[Serial Mgmt]
[Enet Mgmt 10/100]

MP4, MP8
T5-4 or T5-8
Midplane

Memory Processor Module Block Diagram (T5-8 and T5-4)


Xcede HD 6-Row Midplane Connector

IOS
0

IOS
1

CLR
0

CLR
1

CLR
2

CLR
3

CLR
4

CLR
5

CLR
6

IOS
0

IOS
1

CLR
0

CLR
1

T5 (CM1)
MCU0
L0
FSR0

L1
FSR1

MCU1
L0
FSR2

L1
FSR3

CLR
3

CLR
4

CLR
5

CLR
6

T5 (CM0)

MCU2
L0
FSR4

CLR
2

L1
FSR5

MCU3
L0
FSR6

L1
FSR7

MCU0
L0
FSR0

L1
FSR1

MCU1
L0
FSR2

L1
FSR3

MCU2
L0
FSR4

L1
FSR5

MCU3
L0
FSR6

L1
FSR7

BoB1

BoB3

BoB6

BoB5

BoB3

BoB0

BoB5

BoB7

C C
0 1

C C
1 0

C C
0 1

C C
1 0

C C
0 1

C C
1 0

C C
0 1

C C
1 0

BoB0

BoB2

BoB4

BoB7

BoB2

BoB1

BoB4

BoB6

C C
0 1

C C
1 0

C C
0 1

C C
1 0

C C
0 1

C C
1 0

C C
0 1

C C
1 0

CMO / MCU0 / L0 = FSR0 = BOB0


CMO / MCU0 / L1 = FSR1 = BOB1
CMO / MCU1 / L0 = FSR2 = BOB2
CMO / MCU1 / L1 = FSR3 = BOB3
CMO / MCU2 / L0 = FSR4 = BOB4
CMO / MCU2 / L1 = FSR5 = BOB5

25Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

CMO / MCU0 / L0 = FSR0 = BOB0


CMO / MCU0 / L1 = FSR1 = BOB1
CMO / MCU1 / L0 = FSR2 = BOB2
CMO / MCU1 / L1 = FSR3 = BOB3
CMO / MCU2 / L0 = FSR4 = BOB4
CMO / MCU2 / L1 = FSR5 = BOB5

T5 PCIe Subsystem
Dual x8 PCI Express Gen 3 ports provide 32 GB/s peak b/w
Supports Atomic Fetch-and-Add, Unconditional-Swap and
Compare-and-Swap operations
Accelerates virtualized I/O with Oracle Solaris VMs
128k virtual function address spaces ensure direct SR-IOV access for all
logical domains
64-bit DVMA space reduces IO mapping overhead, improving network
performance
Guarantees fault and performance isolation among guest OS instances

Supports PCI Express Power Management


26Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5 PCIe Progression
T4

T5

PCI Express revision

Gen 2 (dual x8 ports)

Gen 3 (dual x8 ports)

Throughput full duplex

16 GBs

32 GBs

Data Management Unit

Single shared unit for both x8


PCIe ports

Two independent units one for


each x8 PCIe port

Physical Address Support

44 bit

48 bit

Transaction Id Identification
on MSI and MSI-X

No

Yes

PCIe Atomic Transactions

No

Yes

TLP Processing Hints

No

Yes, directs data to L3 cache

PCIe 2.0 compliance (ECN


Internal Error Reporting)

Signaled via MSI interrupt

Signaled via PCIe message

27Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5 and M5 PCIe Carrier Card


Supports standard low-profile PCIe cards

Air Flow

PCIe Retimer
x16 Connector
(x8 electrical)
28Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

About the F40 Flash Card (Aura2)


Supported on T5-2, T5-4, T5-8, and M5-32
Available on T5-2 and M5-32 at RR
Available on T5-4 and T5-8 at RR + 1Q
Special carrier card for T5-4 and T5-8 that will span 2 PCIe slots.
Required due to thermal characteristics.
Hot-plug support on T5-4, T5-8, and M5-32 at RR + 1Q
Can not be placed in Slot 8 of IOU on M5-32 due to thermal issues.

29Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5-4: 1 PM PCIe Connectivity


PM0

7
Config
ID

Native Config
(T5-4 only)

PM1

CPU0

CPU1

PFM

MP

Front USB x1
Debug Slot x1

x8 up.
Part. 0

Switch 6

SAS0 x8

Switch 5

NET1 x8

x8 up.
Part. 0

x8 up.
Part. 0

x8 up.
Part. 1

x8 up.
Part. 0

Switch 0

x8 up.
Part. 0

Switch 1

x8 up.
Part. 0

Switch 2

x8 up.
Part. 0

Switch 3

x8 up.
Part. 1

Switch 4

SP VGA x1
Rear USB x1
SAS1 x8
NET0 x8

x8 up.
Part. 1

Debug Slot x1

MP
EB

Dotted line
devices
reside
on RIO

PCI-Express
Low Profile
Hot Plug Slots
Slot #
8 lanes
CPU #

10

11

12

13

14

15

16

x8
c0

x8
c0

x8
c0

x8
c0

x8
c0

x8
c0

x8
c1

x8
c1

x8
c1

x8
c1

x8
c1

x8
c1

x8
c1

x8
c1

x8
xx

x8
xx

T5-4/8 Native 2-Socket Configuration with One Root Domain Block fill color identifies Root Domain ownership
Block outline color identifies association to PM
Single non-redundant Domain
Switch 2 Slots crossed to maintain consistent Slot population order
Second level Switch 6 is partitioned
Slots drawn in order from left to right as in the actual chassis
differently from other configs

30Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5-4: 2 PM PCIe Connectivity


PM0

Native Config

Config
ID

PM1

CPU0

CPU1

CPU2

CPU3

MP

Front USB x1
Debug Slot x1

x8 up.
Part. 0

Switch 6

SAS0 x8

Switch 5

NET1 x8

x8 up.
Part. 0

x8 up.
Part. 0

x8 up.
Part. 1

x8 up.
Part. 0

Switch 0

x8 up.
Part. 3

x8 up.
Part. 0

Switch 1

x8 up.
Part. 3

x8 up.
Part. 0

Switch 2

x8 up.
Part. 3

x8 up.
Part. 1

Switch 3

Switch 4

SP VGA x1
Rear USB x1
SAS1 x8
NET0 x8

x8 up.
Part. 1

Debug Slot x1

MP
EB

Dotted line
devices
reside
on RIO

PCI-Express
Low Profile
Hot Plug Slots
Slot #
8 lanes
CPU #

10

11

12

13

14

15

16

x8
c0

x8
c0

x8
c0

x8
c0

x8
c2

x8
c2

x8
c2

x8
c2

x8
c1

x8
c1

x8
c1

x8
c1

x8
c3

x8
c3

x8
c3

x8
c3

T5-4/8 Native 4-Socket Configuration with One Root Domain Block fill color identifies Root Domain ownership
Block outline color identifies association to PM
Single non-redundant Domain
Switch 2 Slots crossed to maintain consistent Slot population order
Second level Switch 6 is partitioned
Slots drawn in order from left to right as in the actual chassis
differently from other configs

31Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5-8: 2 PM PCIe Connectivity


PM0

Native Config

Config
ID

PM1

CPU0

CPU1

PM2

PFM

PM3

PFM

CPU6

CPU7

MP

Front USB x1
Debug Slot x1

x8 up.
Part. 0

Switch 6

SAS0 x8

Switch 5

NET1 x8

x8 up.
Part. 0

x8 up.
Part. 0

x8 up.
Part. 1

x8 up.
Part. 0

Switch 0

x8 up.
Part. 3

x8 up.
Part. 0

Switch 1

x8 up.
Part. 3

x8 up.
Part. 0

Switch 2

x8 up.
Part. 3

x8 up.
Part. 1

Switch 3

Switch 4

SP VGA x1
Rear USB x1
SAS1 x8
NET0 x8

x8 up.
Part. 1

Debug Slot x1

MP
EB

Dotted line
devices
reside
on RIO

PCI-Express
Low Profile
Hot Plug Slots
Slot #
8 lanes
CPU #

10

11

12

13

14

15

16

x8
c0

x8
c0

x8
c0

x8
c0

x8
c6

x8
c6

x8
c6

x8
c6

x8
c1

x8
c1

x8
c1

x8
c1

x8
c7

x8
c7

x8
c7

x8
c7

T5-8 Native 4-Socket Configuration with Two Root Domains


Two slot per Root Complex
Two Redundant Path Root Domains

32Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Block fill color identifies Root Domain ownership


Block outline color identifies association to PM
Switch 2 Slots crossed to maintain consistent Slot population order
Slots drawn in order from left to right as in the actual chassis

T5-8: 3 PM PCIe Connectivity


PM0

Native Config

Config
ID

PM1

PM2

CPU0

CPU1

CPU2

CPU3

PM3

PFM

CPU6

CPU7

MP

Front USB x1
Debug Slot x1

x8 up.
Part. 0

Switch 6

SAS0 x8

Switch 5

NET1 x8

x8 up.
Part. 0

x8 up.
Part. 0

x8 up.
Part. 1

x8 up.
Part. 1

x8 up.
Part. 0

Switch 0

x8 up.
Part. 1

x8 up.
Part. 3

x8 up.
Part. 0

Switch 1

x8 up.
Part. 2

x8 up.
Part. 3

x8 up.
Part. 0

Switch 2

x8 up.
Part. 2

x8 up.
Part. 3

Switch 3

x8 up.
Part. 1

Switch 4

SP VGA x1
Rear USB x1
SAS1 x8
NET0 x8

x8 up.
Part. 1

Debug Slot x1

MP
EB

Dotted line
devices
reside
on RIO

PCI-Express
Low Profile
Hot Plug Slots
Slot #
8 lanes
CPU #

10

11

12

13

14

15

16

x8
c0

x8
c2

x8
c0

x8
c2

x8
c6

x8
c6

x8
c3

x8
c6

x8
c1

x8
c1

x8
c1

x8
c1

x8
c3

x8
c7

x8
c7

x8
c7

T5-8 Native 6-Socket Configuration with Two Root Domains


Two slot per Root Complex
Two Redundant Path Root Domains

33Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Block fill color identifies Root Domain ownership


Block outline color identifies association to PM
Switch 2 Slots crossed to maintain consistent Slot population order
Slots drawn in order from left to right as in the actual chassis

T5-8: 4 PM PCIe Connectivity


PM0

Native Config 0

Config
ID

PM1

PM2

PM3

CPU0

CPU1

CPU2

CPU3

CPU4

CPU5

CPU6

CPU7

MP

Front USB x1
Debug Slot x1

x8 up.
Part. 0

Switch 6

SAS0 x8

Switch 5

NET1 x8

x8 up.
Part. 0

x8 up.
Part. 0

x8 up.
Part. 1

x8 up.
Part. 1

x8 up.
Part. 0

Switch 0

x8 up.
Part. 1

x8 up.
Part. 2

x8 up.
Part. 3

x8 up.
Part. 0

Switch 1

x8 up.
Part. 1

x8 up.
Part. 2

x8 up.
Part. 3

x8 up.
Part. 0

Switch 2

x8 up.
Part. 1

x8 up.
Part. 2

x8 up.
Part. 3

Switch 3

x8 up.
Part. 0

x8 up.
Part. 1

Switch 4

SP VGA x1
Rear USB x1
SAS1 x8
NET0 x8

x8 up.
Part. 1

Debug Slot x1

MP
EB

Dotted line
devices
reside
on RIO

PCI-Express
Low Profile
Hot Plug Slots
Slot #
8 lanes
CPU #

10

11

12

13

14

15

16

x8
c0

x8
c2

x8
c0

x8
c2

x8
c4

x8
c6

x8
c4

x8
c6

x8
c1

x8
c3

x8
c1

x8
c3

x8
c5

x8
c7

x8
c5

x8
c7

T5-8 Native 8-Socket Configuration with Two Root Domains


Two slot per Root Complex
Two Redundant Path Root Domains

34Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Block fill color identifies Root Domain ownership


Block outline color identifies association to PM
Switch 2 Slots crossed to maintain consistent Slot population order
Slots drawn in order from left to right as in the actual chassis

PCIe Data Paths: Full System


Two root complexes per T5
processor
Each PCIe port on a T5 processor
controls a single PCIe slot

35Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Root Complex

Each processors root complexes


must be assigned to the same
domain
Example:
/pci@300/pci@1/pci@0/pci@6
& /pci@340/pci@1/pci@0/pci@6
from CPU0 are assigned to the
same logical domain

36Copyright 2012, Oracle and/or its affiliates. All rights reserved.

PM

CPU

Switch

I/O Slot

/pci@300/pci@1/pci@0/pci@6

/pci@340/pci@1/pci@0/pci@6

/pci@380/pci@1/pci@0/pci@a

11

/pci@3c0/pci@1/pci@0/pci@e

/pci@400/pci@1/pci@0/pci@c

/pci@440/pci@1/pci@0/pci@c

10

/pci@480/pci@1/pci@0/pci@4

12

/pci@4c0/pci@1/pci@0/pci@8

/pci@500/pci@1/pci@0/pci@e

/pci@540/pci@1/pci@0/pci@e

13

/pci@580/pci@1/pci@0/pci@a

15

/pci@5c0/pci@1/pci@0/pci@8

/pci@600/pci@1/pci@0/pci@8

/pci@640/pci@1/pci@0/pci@8

14

/pci@680/pci@1/pci@0/pci@4

16

/pci@6c0/pci@1/pci@0/pci@4

Confidential Oracle Internal

Root Complex Path

Zakim (ZK) M5/T5 Systems Memory Interface


Also referred to as a BoB (Buffer-on-Board)
Features and Technology:
4 ZKs per M5 memory controller
2 ZKs per T5 memory controller
Supports DDR3 DIMMs
Memory Link to DDR Interface
2 Memory Link Ports, 2 DDR ports
8 Write FIFOs
Pass through commands to DIMMs
Core frequency: 1066 MHz

37Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5-4 and T5-8 Memory Config Rules

All DIMMs must have the same Oracle Part Number on a PM (Processor
Module).
All PM's must be fully populated (32 DIMMs)

No partial memory configs available at RR


All PM's in the system must be configured the same (ATO specific rule)

38Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5-4 and T5-8 Airflow/Cooling


T5-8
(top 3U)

T5-4
(5U)

39Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

SPARC T5 Disk Drives

Disk drives are 2.5 form factor


Chassis supports SAS HDD and SATA SSDs
SAS-2 HDDs

300GB @ 10K RPM


600GB @ 10K RPM
900GB @ 10K RPM

SATA SSDs

100GB
300GB
Disk LED's

Ready to
Remove
Fault
Status

40Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

SPARC T5 Disk Controller


T5 has Dual LSI SAS2008 8port SAS3/SATA2 controllers
Support for RAID 0 (striping) and RAID 1 (mirroring) using 'raidctl'

41Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

The SPARC T5 Processor

42Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

SPARC T5 Processor

Features
16 S3 cores, 16-128 Strands
@ 3.6Ghz

Single or multi-threaded
operation per core

System scalability to 8
sockets
SPARC Core S3

1-8 Strand Dynamically


Threaded Pipeline

ISA-based Cryptoacceleration
8MB Shared L3$

43Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Integrated I/O

Double I/O bandwidth over T4

2 x8 Lane PCIe 3.0 @ 8GT/s


System Scalability

7 Coherence Ports for scalability


to 8S
Power Management

Dynamic Voltage Frequency


Scaling

Downclock, Overclock

DDR3 1066 MHz

BoB

DDR3 1066 MHz

BoB

BoB

DDR3 1066 MHz

BoB

BoB

BoB

DDR3 1066 MHz

BoB

BoB

Memory
Control

Memory
Control

Memory
Control

Memory
Control

Coherence Unit

Coherence Unit

Coherence Unit

Coherence Unit

SPARC S3
Core

L3$ B0
L3$
B0
1MB,1
6-way

L3$ B2
1MB,1
6-way

L3$ B1
1MB,1
6-way

L3$ B3
1MB,1
6-way

L3$ B4
B0
L3$
B0
1MB,1
6-way

L3$ B6
1MB,1
6-way

L3$ B5
1MB,1
6-way

L3$ B7
1MB,1
6-way

8 x 9 Crossbar (~1TBps bandwidth)

FGU
Crypto

Coherency Links
12.8 Gbps per lane
- 12 lanes per link
Link 0

C0

C1

C2

C3

C4

C5

C6

8 threads per Core


44Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

C7

C8

C9

C10 C11 C12 C13 C14 C15

Coherency 4x4 Switch

16 KB L1D$

IO
Subsystem

Link 1

128 KB L2$
16 KB L1I$

2 x 8 PCIe 3.0 @ 8 GBps


16 GBps each direction

SPARC T5 CPU Block Diagram

Link 2
Link 3

Link 4
Link 5

Link 6

T5 Processor Overview
SerDes

MI/O

SPAR SPAR
C
C
Core Core
SPAR SPAR
C
C
Core Core
MC MC
U
U
SPAR SPAR
C
C
Core Core
SPAR SPAR
C
C
Core Core

SerDes
Confidential Oracle Internal

8MB shared L3 Cache


8 DDR3 BL8 Schedulers

providing 80 GB/s BW

SerDes

SerDes

SPAR SPAR Pwr PCIe


C
C
Gen3
Core Core
SPAR SPAR
C
C
L3 L3 L3 L3
Core Core
Cross
MC MC
U
U
Bar
SPAR SPAR L3 L3 L3 L3
C
C
Core Core
SPAR SPAR
C
C
Core Core Coherence

45Copyright 2012, Oracle and/or its affiliates. All rights reserved.

16 S3 cores @ 3.6GHz

SerDes

8-way 1-hop glueless

scalability
Integrated 2x8 PCIe Gen 3
Advanced Power Management

with DVFS

SPARC T5: Processor Key Features


SPARC T5 based on SPARC S3 core: Same core as used in T4
16 S3 cores, dual PCI Express 3.0 root complexes
Up to eight T5 processors per system, 16 cores x 8 threads 128 (on T5-1B) or up to 1024
threads (on T5-8)

Clock frequency is 3.6 GHz


Each SPARC S3 core contains:

2 Integer pipelines

1 FGU pipeline (consisting of 3 physical sub-pipelines):


FPX pipeline
FGX pipeline
FPD pipeline
1 Load-Store (Memory) pipeline

46Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

SPARC T5: Processor Key Features


Each Core has two Level 1 cache memories, one for data and one for
instructions, each 16KB in size
Each Core has a Level 2 unified cache, 128KB in size

Caches are all inclusive: L3 inclusive of L2; L2 inclusive of L1 (in this context,
inclusive refers to the fact a cached entry is always present in the next higher
level of cache)
Each core on SPARC T5 is capable of OoO execution, dual-issue of
instructions but in order commit.
Each core on SPARC T5 also includes cryptographic acceleration hardware,
accessible via user-level instructions.

47Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

SPARC T5: Processor Key Features


Memory systems based upon SPARC T5 use DDR3 Quad Rank DIMMs
DIMMs used are Registered DIMMS (RDIMMs)
16 DIMM slots per Processor memory capacity
depends on size of DIMM chosen for system
8 GB, 16 GB DIMM
ECC DIMMs required to be used at 1066 MT/s
Newly designed BoB between MCU and DIMMs
Four MCUs per processor; each maintains a memory link speed of 12.8 Gbps
Protocol between each MCU and its two BoBs is proprietary in nature
48Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

S3 Core Recap
28nm port from 40nm T4
Out-of-order, dual-issue
High frequency achieved with 3.6GHz
16 stage integer pipeline
Dynamically threaded, one to eight strands
Accelerates 16 encryption algorithms and random number generation

49Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

SPARC T4/T5/M5 Leads in On-Chip Encryption Acceleration


Processor / Mechanisms

SPARC T4, T5, and M5

IBM Power7

IBM Power7+

Intel Westmere/
Sandybridge

Operational Model

Userland unprivileged access


to on-core cryptographic
functions

none

3 accelerators shared
across 8 cores

Userland

Asymmetric /Public Key


Encryption

RSA, DH, DSA, ECC

none

RSA, ECC

RSA, ECC

Symmetric Key / Bulk Encryption

AES, DES, 3DES,


Camellia, Kasumi

none

AES

AES

Message Digest /
Hash Functions

CRC32c, MD5, Sha-1,


SHA-224, SHA-256, SHA384, SHA-512

none

MD5, SHA-1, SHA-256,


SHA-512

none

Random Number Generation

Supported

none

Supported

none

API Support

PKCS#11, Ucrypto APIs, JCE

none

PKCS#11

Intel IPP libraries

Virtualization Support

Solaris Zones
Oracle VM for SPARC

none

??

Intel VT

Crypto acceleration now available from 1 processors systems up to 32-way systems!


50Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

SPARC Core Roadmap


S5 Core
M5 Servers
T4 Servers
S3 Core
T3 Servers
S2 Core
M-Series
SPARC64 VII/VII+ Core

51Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5 Servers

M6 Servers

S3 Core Overview
8-way threaded, dual-issue, OoO execution, in order commit
Dynamically threaded with hardware-optimized resource sharing
Support for Critical Threads
Deep pipeline for high frequency operation (3 GHz in 40 nm)
Balanced single-thread and multi-thread performance
5X better single-thread than SPARC T3 with equivalent multi-thread
performance
Enhanced instruction set to accelerate Oracle SW stack
PAUSE, fused compare-branch
Integrated user-level cryptographic acceleration
DES/3DES, AES, Kasumi, Camellia, MD5, SHA-1, SHA224/256/384/512,RSA, DSA, CRC32c
Foundation core for future technology / product nodes
52Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Oracle SPARC S3 Core


Enables T4, T5, M5

Design Goals
Develop a common replacement core for T-series processors
Significantly improve the single strand performance of the T3 processor
-

Significantly improve the throughput performance of the M3 processor

Improve the RAS and power management capabilities


Maintain backward ISA, Solaris and OVM for SPARC compatibility

Frequency
L1 Instruction Cache
L1 Data Cache
L2 Cache
# of Pipelines
# of Threads per Pipeline
Instructions per Thread
Out of Order Issue
Cryptography Acceleration
OVM for SPARC Compatible
SPARC V9 ISA Compatible

53Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Core S1: M2, M3

Core S2: T2, T2+, T3

Core S3: T4, T5, M5

2.4 3.0 GHz

1.4 1.65 GHz

64KB
64KB
1
2

16KB
8KB
2
4

T4: 2.85 3.0 GHz


T5, M5: 3.6 GHz
16KB
16KB
128KB
1
1-8 Dynamic

4 per cycle
Yes
None
No
Yes

1 per cycle
No
SPU
Yes
Yes

2 per cycle
Yes (36 instr window)
ISA Based
Yes
Yes

Confidential Oracle Internal

S3 Core: Dynamic Threading


While software can activate up to 8 strands on each core at a time, hardware
dynamically and seamlessly allocates core resources such as instruction, data,
and L2 caches and TLBs, and out-of-order execution resources such as the 128entry re-order buffer in the core among the active strands.
Software activates strands by sending an interrupt to a HALTed strand. Software
deactivates strands by executing a HALT instruction on each strand it wants to
deactivate. No strand has special hardware characteristics; all strands have
identical hardware capabilities.

54Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

S3 Core: Dynamic Threading


Since the core dynamically allocates resources among the active strands, there
is no explicit "single-thread mode" or "multi-thread mode" for software to activate
or deactivate.
If software effectively halts all strands except one on a core via Critical Thread
Optimization, the core devotes all of its resources to the sole running strand.
Thus, that strand will run as quickly as possible

55Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

S3 Core: Critical Thread Optimization


The S3 core, starting with Oracle Solaris 10 8/11, can optimize performance by
assigning one software thread exclusive access to all of a core's hardware
resources. That software thread is considered to be a "critical thread."
Solaris automatically detects opportunities to perform this assignment: when one
software thread has high CPU utilization and there are more cores than runnable
threads. We recommend that users allow Solaris to automatically perform Critical
Thread assignment.
A privileged user can tell Solaris that a particular software thread should be a
critical thread, via the nice(1) command. Solaris will then assign that thread to a
core, even if there are more runnable threads than cores.

56Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

S3 Core: Critical Thread Optimization


Solaris Critical Threads optimization for S3 core, tries to provide exclusive access
of certain hardware resources to certain application threads
Since the core dynamically allocates resources among the active strands, there
is no explicit "single-thread mode" or "multi-thread mode" for software to activate
or deactivate
There is no new API for declaring threads as critical; that would require
significant changes to source code
Rather, to invoke Critical Thread Optimization, use the following CLI or system
calls to flag a thread as critical by raising its priority to 60:
priocntl(1)
priocntl(2)
priocntlset(2)
57Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

S3 Core: Critical Thread Optimization


Starting with Oracle Solaris 10 8/11, a thread is declared to be critical if raised to
priority 60; the thread can be in any scheduling class
In Oracle Solaris 11, to be considered critical by the scheduler, a thread must be:
in the FX (Fixed Priority) or RT (Real-Time) scheduling classes
be raised to priority 60 by one of the previously mentioned mechanisms
In either of the above instances, this one thread will run as quickly as possible as
it has exclusive access to all core resources

58Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Critical Threads for key applications


Database

Applicability

Opportunity

Current Status

Logwriter, LMS

Up to 30% improvement in
efficiency

LMS is already CT ready.


LGWR planned for 12c

JAVA (JVM)

Compiler threads, GC Up to 2x improvement for app Support for JVM and JAVA
and priority mapping
apps to be CT aware is
startup,
Smooth
GC
support
integrated in JDK7U4

Coherence

Packet writer, service


thread

Solaris

S11U1 / S10U11

59Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Up to 20% improvement in
throughput

Integrated in Coherence
version 3.7.1 Patch 1

Improve CT perf to be within Optimizations for decayed


10% of best case (hand
PG util and stealing being
optimized)
integrated in S11U1

Internode Coherency Overview


Glueless 1-hop scaling to eight sockets

T5

T5

Glueless means no hub required to connect all 8


CPUs

striped across all processors


stored in on-chip SRAMs
flexible for different socket counts

Higher BW efficiency than snoop-based


protocols enables better scaling
50% more effective bandwidth than
comparable snoopy implementation
60Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5

T5

T5

T5

T
-

A precise directory tracks all L3s in the


system

T5

T5

Internode Performance Optimizations


Speculative memory reads prior to cache line serialization in the directory
Cache-to-cache line transfers between nodes
Dynamic congestion avoidance routes inter-node data around congested links

61Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5 System Interconnects

2-Way
Dual Socket

1-Way
Single Socket

4-Way

6-Way
8-Way
62Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

8 Socket Local Coherency & Data Interconnect


All-to-All Interconnect
Where the Node-to-Node Fabric is 12 diff
pairs per link in each direction.
DIMMS
T5 interconnect bandwidth= 157.5 GB/sec
T5-8 interconnect bandwidth= 1260 GB/sec

DIMMS

M5/T5

M5/T5

PCIe Gen3 Bandwidth


8 diff pairs per ports
At 8Gb/sec ~8GB/sec/direction
~8Gb/sec/lane X 8 lanes = 64 Gb/s
X 2 directions
X 2 Ports/chip
DIMMS
= 256Gb/s/chip
= 32 GB/s per chip
DDR3-1066
Memory Bandwidth
T5 is 133 GB/sec
T5-8 is 1064 GB/s

63Copyright 2012, Oracle and/or its affiliates. All rights reserved.

M5/T5

DIMMS

M5/T5

DIMMS

M5/T5

DIMMS

POINT-TO-POINT
LOCAL
INTERCONNECT
M5/T5

DIMMS

Confidential Oracle Internal

M5/T5

M5/T5

DIMMS

Latency for T5

T5-2

T5-4

Local Memory

136ns

Remote Memory

209ns

Cache to Cache

64Copyright 2012, Oracle and/or its affiliates. All rights reserved.

127ns

Confidential Oracle Internal

146ns

T5-8

155ns

Scalability of T4 vs T5
By numbers

Feature

Link Bandwidth

T4 Snoopy Based
Coherence Protocol
8 node snoops will consume 25% Link B/W &
increases linearly w/more nodes
Address serialization is done at Home Node. Home
Node broadcasts snoop request to all nodes. All
nodes except the requesting node require to
participate the snoop operation and provide snoop
response back to requesting node.
Message broadcast and response consume a lot of
link bandwidth.

T5 Directory Based
Coherence Protocol
8 node directory based will consume 5% of Link B/W
Address serialization is done at Directory Node.
Directory Node keeps track of which node hold each
cache line. Eliminating the need for broadcasting,
and relieve the L3$ from unnecessary foreign snoop
operation.
Directory filter the snoops sent to the share nodes.
Allow link bandwidth to be used more efficiently.

L3$ need to participate every snoop request from any Only the L3$ from the selective node require to
L3$ Performance other node. The L3$ performance can be dropped
participate the foreign snoop operation. L3$ has less
due to lots of foreign snoop requests.
distraction from foreign snoop request.
Easy to scale to large number of processor
Scalability
Limited to small scale of system.
environment.

65Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Memory Controller of T4 vs T5
T4 MCU

T5 MCU

6.4 Gb/s

12.8 Gb/s

Legacy Intel FBDIMM2 Protocol

Advanced In-house Link Protocol

Not Supported

L0s, L1

Memory Buffer

Intel Milbrook2 MB

Advanced In-house MB

DDR3 Protocol

Burst length of 4

Burst length of 8

DDR3 Speed

800/1066

1066

DDR3 Device

1Gb/2Gb

2Gb/4Gb

Memory Link Speed


Memory Link Protocol
Memory Link Low Power
Feature

66Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

RAS

67Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Definition of Terms

Hot-plug:
refers to the fact that a component can be plugged and unplugged without
powering down the platform. It applies to both hot swap and hot service.
Hot service:

refers to the ability to perform hot-plug operations, with the additional


necessity of some operator actions (invocation of a CLI or actuating a hot
service button on the component to be removed).

The system will notify the user when it is safe to remove the component.

Typical examples would be PCIe Express modules.


Hot swap:

refers to an operation where a component is unplugged and plugged in with


no interaction with the ILOM or domain required.

Typical examples here are a single RAID disk or a power supplies.

68Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

SPARC T5 System RAS Overview

Designed to minimize part count and operating temperature to enhance


reliability
End-to-end data protection detecting and correcting errors throughout server
ECC everywhere
Processor and Memory protection

Major components redundant & hot-pluggable

CPU core and thread off-lining


Memory with ECC, x4/x8 DRAM Extended ECC, page retirement, and lane
failover
Fan, Power Supply, and internal disks
RAID capability for internal disks

Fault Management Architecture (FMA) support on ILOM

69Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

End-to-End RAS
Built RAS from the inside out
Start with the processor, then memory,
system and IO, virtualization layer, and the
OS
Add Oracle Solaris Cluster software for
additional service availability
Fault Management Architecture (FMA)
binds all the layers together

70Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

T5/M5 Systems RAS


S11 FMA
Diagnosis engine on SP
Auto reconfigure on failure
Soft Error Rate Discrimination (SERD)
Bad page retirement
OS and SP watchdogs
FMA Component hot-upgradeable

Hypervisor
Enables software partitioning (LDoms)
virtualization and failure containment
Processor support for error clearing, correction
and collection

T5/M5 Processor
L1$ Tag, Status $ Data
Parity protection
Retry on error
L2$/L3$ Data
SEC/DED protection
Cache-line Sparing
L2$/L3$ Tags
SEC/DED protection
Inline Correction
Cache-line Sparing
L2$/L3$ Status & Directory
SEC/DED protection
Central Directory and Switch
Inline Correction
SEC/DED protection with in line correction
Architectural RegistersL2 Cache
Physical domain isolation
SEC/DED protection
CRC protected System Interconnect with message retry and
Precise Trap and
lane sparing
Hypervisor Correction and Retry
Deconfigurable directory chips, no loss of functionality,
minimized bandwidth loss
Unique to M5
Redundant Scalability Switch Boards

71Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

System
Redundant SPs with automatic failover
Redundant clock boards
Diagnosis to the FRU level on first fault
Power and Cooling
Advanced Power Management
Redundant hot-swap fans
Redundant hot-swap AC/DC
Dual grid power
System I/O
PCI-Express end-to-end CRC
PCI Express link retry
Hot-plug low profile PCI Express cards
Redundant, hot-plug boot disks
Alternate connections between M5 and IO
controllers
Memory
SDRAM Soft Errors
ECC Protection and Correction
Extended ECC Protection
4-bit Correction
Pin Steering
Channel Interconnect
CRC protection/Message Retry
Lane Sparing

SPARC RAS Features


RAS Features

T4-1/T4-2

T4-4

M4000/M5000

T5-2

T5-4/T5-8

M8000/M9000

M5-32

Dynamic Reconfiguration

Processor L2$ Degradation

Processor L1$ Degradation

L3$ Line Sparing

X
X

Hot-plug PCIe

Memory Mirroring

X
X

Memory Lane Sparing

CPU interconnect Lane Sparing

Hard Domains

Common
Features
Logical Domains

Hot-swap power supplies, fans

Hot-plug disks
X
X

ECC Memory

Redundant System Interconnect


72Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Extended-ECC Memory

Instruction Retry
X
X
Partial ASIC Recovery
X

Confidential Oracle Internal

X
X

Fault Management

Knowledge Articles in MOS

ILOM fdd Diagnosis

Faults and Alerts

No ALOM Compatibility

ILOM FMA Captive Shell

Sideband Service Processor Network Connection

New ILOM Fault Notification (SNMP Trap)

ASR Support

FMA on M5 ILOM also applies to T5 ILOM, except for M5 specific features

73Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

FMA restricted shell in SPSH


To get FMA details in SPSH
start -script /SP/faultmgmt/shell
Returns a faultmgmtsp> prompt
Available build-in commands:
echo - Display information to user.

Typical use: echo $?


help
Exit exits restricted shell
External commands:
fmadm - Administers the fault management service
fmdump - Displays contents of the fault and ereport/error logs
fmstat - Displays statistics on fault management operations
etcd - ereport injector
74Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Fault Management on T5 systems

T5 CPU and Memory faults are now diagnosed by ILOM

Disabled Database (DDB) owned by ILOM

FMA's Fault Proxy is used to keep ILOM's fault manager in sync with Solaris' fault
manager. Both will display the sum of all faults in the system.
Faults can be repaired from either side.
Fault Proxy communicates via the Ethernet Over USB connection.
IO faults are still diagnosed by Solaris.
For faults which diagnose resources as unusable, ILOM will add those resources to
the DDB. Resources excluded on next host reset.
When faults are repaired, ILOM automatically updates the DDB. Bringing
components back online requires a host reset.

Extended SP-POST (Power on Self Test)

Runs at SP boot. Tests devices on the SP FRU and its Ethernet port.
Status stored and converted to ereports after ILOM boots.

75Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Fault proxy
SP

ereports
hostd

FETD
ip-transprt

LDC

Control Domain
ETM

ETM

faults
TCP/IP

ereports
LDC

IO Domain
ETM

faults
ip-transport

ETM

LDC

ETM

IO ereports are forwarded from the SP to the control domain, and then on to any
relevant IO domain
Faults are proxied between the SP, the control domain and any IO domains to provide
a single view of faults in the system.
Non-servicable faults such as memory faults are not proxied.

The SP and the control domain can view and manage all faults in the system.
An IO domain can only view and manage faults local to the domain.

76Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

Degraded HW Configurations T5-8


If a CPU has no configurable memory, i.e., all memory links are unconfigurable, then the CPU itself is
unconfigurable
A T5-8 will not operate with only 5 or 7 CPUs configured
One more CPU(s) must be chosen to be deconfigured
For an 8-way, if we fault a CPU, we will offline the other CPU on the same PM
For a 6-way, if we fault a CPU, we will also offline the other CPU on the same PM

77Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

ASR Support

SPARC T5 servers will be supported by ASR (Automatic Service Request) at


release

Continues use of sunHwTrapFaultDiagnosed SNMP notification

Telemetry for ILOM fdd diagnosis

Supports platform and FRU identity

Supports multi-suspect list

78Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

79Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

80Copyright 2012, Oracle and/or its affiliates. All rights reserved.

Confidential Oracle Internal

You might also like