Fujitsu vs All or Japanese Killer RISC Servers
Greetings Colleagues!
I was encouraged to write this article by a conversation with a colleague (a good engineer and a specialist in his field), in which he mentioned that the x86 server market is “exhausted” - all manufacturers copy each other, all servers are the same and no one offers anything new. At the same time, the market of RISC \ UNIX servers, due to its smaller size, has also changed imperceptibly for the mass consumer of servers:
• Oracle, c inherited from SUN SPARC systems, relies on the Exa family built on the x86 platform.
• Intel has long ceased to develop the direction of Itanium
• HP in a court of law forced Oracle to support servers based on Itanium Super DOME .......
• IBM is proactive in the sequential development of the IBM POWER processor family. But still, IBM doesn't do it as intensively as Intel and doubles its performance every three years. The result is a 4-socket Intel x86 machine with Intel Xeon E7 v2 processors ahead of Oracle 10g in OLTP load, similar to the Power750 based on four PowerP7 + by 25-30% ... But for some reason no one throws IBM Power in the trash and does not run headlong behind x86 servers?
The answer is simple - fault tolerance.
Traditionally, it was a growing business that required reducing unplanned downtime and minimal data loss and recovery time. RISC \ Unix servers responded to this task, but at the same time, moving away from the x86 platform required significant financial injections. The servers themselves were more expensive, had an annual mandatory payment for a subscription to a service (the so-called maintenance), had a proprietary architecture closed (each manufacturer has its own). And another set of applications (for example, MS SQL does not work on RISC machines - you need to migrate to DB2 or Oracle) and the need for specially trained people who can manage and maintain such servers ... All this created difficulties and was reflected in the final budget, and many people realized this customers stayed on x86.
Fujitsu went the other way. Using its many years of experience in designing fault-tolerant servers, the company launched its first server in 1958 and delivered the mainframe AMDAHL 470V / 6 in 1978 to the US National Aeronautics and Space Administration (NASA).
In 1995, the company began producing Sparc64 processors and systems based on these processors.
In 2002, the fastest mainframe GS21-600 was released ...
Later there were various models on SPARC processors and the PRIMEQUEST family of machines using x86 architecture and meeting all the requirements for mission-critical servers began to emerge.
The first server, built on the patterns of mainframe servers and including Intel Xeon E7 processors, was called PRIMEQUEST 1000 and was released in 2012.
At the beginning of this year, along with the announcement of new Intel Xeon E7 v2 processors, PRIMEQUEST 2000 was released.
What are its main differences from the usual 4-8 socket servers from other manufacturers?
Another internal architecture is perfect, in short - this is a mainframe with x86 processors. That is, we get the performance of x86 processors along with new features, such as hardware partitions and the highest fault tolerance.
In traditional servers, as measures to ensure fault tolerance:
ECC in the memory subsystem
The presence of two power supplies and fans with "hot" replacement
Active PCI slots
For comparison, the capabilities of PRIMEQUEST and the most common RISC \ UNIX machines are tabulated. Differences from traditional x86 servers are that hot-swappable motherboards are implemented and this implementation is made better than in modern RISC machines.
The hardware partitions allow you to divide the server into several parts, just as modern ESX or Hyper-V virtualization tools do. The main differences are that the hardware-implemented circuit provides complete electrical independence between partitions, does not consume resources and does not require licenses to create partitions. The software licensing scheme is also changing.
The capabilities of the hardware partitions are visible below:
Now we get the fault tolerance implemented inside the server: some hardware problems inside the partition (output of the motherboard, processor, etc.) do not stop the server or other partitions. Advanced sections allow you to divide a specific hardware partition into additional subsections with the allocation of resources with great accuracy.
It is also possible to use with classic software hypervisors such as for example ESX from VMware or Hyper-V from Microsoft.
If you use an application the size of a large enterprise that can only work on an x86 platform, such as SAP HANA, then it gets the opportunity to increase fault tolerance at the hardware level.
Currently, dynamic reconfiguration capabilities are only possible in RedHat version 7.
Support for hardware partitions can significantly save on software licenses. As an example, take the same Oracle Database as the most common software.
As an illustrative example, let's take the tpm OLTP test results Oracle 10g
Power 750, POWER7 +, 4000 MHz, 2 MB SLC (8-core 80M TLC per DCM) - 3,995,403
PRIMEQUEST-2800E, Xeon E7-8890 v2, 2800 MHz, 3.75 MB SLC ( 15-core 37.5M TLC) - 9,400,000
Many will probably notice right away that PRIMEQUEST has 120 cores, and Power 750 only 32. Yes, this is true, and if we take the proportion to the Power of the machine, we get about 48 cores to achieve that same performance. Accordingly, we compare the hardware partition of 48 cores with 32 cores from IBM. But if we recall the licensing rules (remember that the PPAR hardware partitions are recognized by Oracle), the coefficient for Intel Xeon is 0.5, while for Power7 + it is 1. Further, simple math:
In total, the difference is about $ 400,000. Plus, every year you need to pay another 21% of the cost of licenses for subscription - another $ 264,000 in savings for 3 years.
Such partitions are supported by Oracle: if a regular x86 server must be fully licensed, then if PRIMEQUEST is used, only the partition where Oracle is running. Also, if an error occurs on the hardware partition, Oracle Support considers this an error on the physical machine. In the case of using software virtualization (ESX, Hyper-V, etc.), Oracle support will ask you to repeat a similar situation on a physical machine.
Intel also has low-core models, but with a high clock speed. For example, E7-8893 v2.
The results are 25% more than Oracle 10g OLTP - and if expressed in the language of money:
PRIMEQUEST-2800E- 48 core = rate 0.5 * 48 * 49.600 (price per core) = 1,190,400 $ - 25% = $ 297.5k
That is total savings can reach the sum of about 700 to $ for the initial purchase and more than 562k $ for a subscription for 3 years ... In my opinion, a very convincing argument.
Today I tried to highlight the main features of PRIMEQUEST servers, although it’s possible to continue for a long time: the modular design alone (the server uses PCI switches and the input / input subsystem is expanded by external PCI exp-box - up to 56 pieces) deserves close attention, although who is closely acquainted with RISC servers will see an identical approach.
Summarizing, this server is unique at the moment in the market and is ready to significantly change the view on fault tolerance, functionality for x86 platforms and the total cost of ownership.
I was encouraged to write this article by a conversation with a colleague (a good engineer and a specialist in his field), in which he mentioned that the x86 server market is “exhausted” - all manufacturers copy each other, all servers are the same and no one offers anything new. At the same time, the market of RISC \ UNIX servers, due to its smaller size, has also changed imperceptibly for the mass consumer of servers:
• Oracle, c inherited from SUN SPARC systems, relies on the Exa family built on the x86 platform.
• Intel has long ceased to develop the direction of Itanium
• HP in a court of law forced Oracle to support servers based on Itanium Super DOME .......
• IBM is proactive in the sequential development of the IBM POWER processor family. But still, IBM doesn't do it as intensively as Intel and doubles its performance every three years. The result is a 4-socket Intel x86 machine with Intel Xeon E7 v2 processors ahead of Oracle 10g in OLTP load, similar to the Power750 based on four PowerP7 + by 25-30% ... But for some reason no one throws IBM Power in the trash and does not run headlong behind x86 servers?
The answer is simple - fault tolerance.
Traditionally, it was a growing business that required reducing unplanned downtime and minimal data loss and recovery time. RISC \ Unix servers responded to this task, but at the same time, moving away from the x86 platform required significant financial injections. The servers themselves were more expensive, had an annual mandatory payment for a subscription to a service (the so-called maintenance), had a proprietary architecture closed (each manufacturer has its own). And another set of applications (for example, MS SQL does not work on RISC machines - you need to migrate to DB2 or Oracle) and the need for specially trained people who can manage and maintain such servers ... All this created difficulties and was reflected in the final budget, and many people realized this customers stayed on x86.
Fujitsu went the other way. Using its many years of experience in designing fault-tolerant servers, the company launched its first server in 1958 and delivered the mainframe AMDAHL 470V / 6 in 1978 to the US National Aeronautics and Space Administration (NASA).
In 1995, the company began producing Sparc64 processors and systems based on these processors.
In 2002, the fastest mainframe GS21-600 was released ...
Later there were various models on SPARC processors and the PRIMEQUEST family of machines using x86 architecture and meeting all the requirements for mission-critical servers began to emerge.
The first server, built on the patterns of mainframe servers and including Intel Xeon E7 processors, was called PRIMEQUEST 1000 and was released in 2012.
At the beginning of this year, along with the announcement of new Intel Xeon E7 v2 processors, PRIMEQUEST 2000 was released.
What are its main differences from the usual 4-8 socket servers from other manufacturers?
Another internal architecture is perfect, in short - this is a mainframe with x86 processors. That is, we get the performance of x86 processors along with new features, such as hardware partitions and the highest fault tolerance.
In traditional servers, as measures to ensure fault tolerance:
ECC in the memory subsystem
The presence of two power supplies and fans with "hot" replacement
Active PCI slots
For comparison, the capabilities of PRIMEQUEST and the most common RISC \ UNIX machines are tabulated. Differences from traditional x86 servers are that hot-swappable motherboards are implemented and this implementation is made better than in modern RISC machines.
The hardware partitions allow you to divide the server into several parts, just as modern ESX or Hyper-V virtualization tools do. The main differences are that the hardware-implemented circuit provides complete electrical independence between partitions, does not consume resources and does not require licenses to create partitions. The software licensing scheme is also changing.
The capabilities of the hardware partitions are visible below:
Now we get the fault tolerance implemented inside the server: some hardware problems inside the partition (output of the motherboard, processor, etc.) do not stop the server or other partitions. Advanced sections allow you to divide a specific hardware partition into additional subsections with the allocation of resources with great accuracy.
It is also possible to use with classic software hypervisors such as for example ESX from VMware or Hyper-V from Microsoft.
If you use an application the size of a large enterprise that can only work on an x86 platform, such as SAP HANA, then it gets the opportunity to increase fault tolerance at the hardware level.
Currently, dynamic reconfiguration capabilities are only possible in RedHat version 7.
Support for hardware partitions can significantly save on software licenses. As an example, take the same Oracle Database as the most common software.
As an illustrative example, let's take the tpm OLTP test results Oracle 10g
Power 750, POWER7 +, 4000 MHz, 2 MB SLC (8-core 80M TLC per DCM) - 3,995,403
PRIMEQUEST-2800E, Xeon E7-8890 v2, 2800 MHz, 3.75 MB SLC ( 15-core 37.5M TLC) - 9,400,000
Many will probably notice right away that PRIMEQUEST has 120 cores, and Power 750 only 32. Yes, this is true, and if we take the proportion to the Power of the machine, we get about 48 cores to achieve that same performance. Accordingly, we compare the hardware partition of 48 cores with 32 cores from IBM. But if we recall the licensing rules (remember that the PPAR hardware partitions are recognized by Oracle), the coefficient for Intel Xeon is 0.5, while for Power7 + it is 1. Further, simple math:
In total, the difference is about $ 400,000. Plus, every year you need to pay another 21% of the cost of licenses for subscription - another $ 264,000 in savings for 3 years.
Such partitions are supported by Oracle: if a regular x86 server must be fully licensed, then if PRIMEQUEST is used, only the partition where Oracle is running. Also, if an error occurs on the hardware partition, Oracle Support considers this an error on the physical machine. In the case of using software virtualization (ESX, Hyper-V, etc.), Oracle support will ask you to repeat a similar situation on a physical machine.
Intel also has low-core models, but with a high clock speed. For example, E7-8893 v2.
The results are 25% more than Oracle 10g OLTP - and if expressed in the language of money:
PRIMEQUEST-2800E- 48 core = rate 0.5 * 48 * 49.600 (price per core) = 1,190,400 $ - 25% = $ 297.5k
That is total savings can reach the sum of about 700 to $ for the initial purchase and more than 562k $ for a subscription for 3 years ... In my opinion, a very convincing argument.
Today I tried to highlight the main features of PRIMEQUEST servers, although it’s possible to continue for a long time: the modular design alone (the server uses PCI switches and the input / input subsystem is expanded by external PCI exp-box - up to 56 pieces) deserves close attention, although who is closely acquainted with RISC servers will see an identical approach.
Summarizing, this server is unique at the moment in the market and is ready to significantly change the view on fault tolerance, functionality for x86 platforms and the total cost of ownership.