Servers

Scaling up vs. scaling out: Which makes more sense?

The day-to-day infrastructure administrator will inevitably come to a decision point on whether to scale up or scale out when it comes to servers and systems. IT pro Rick Vanover shares his opinions on one of today’s dividing issues.

A few years ago, we were frequently pressed with decisions such as whether or not to select two or four processor (socket) systems. Processor architectures have continually increased the number of cores available, and in many situations IT administrators could choose the smaller two-processor system in favor of loading up on cores. This was one of my first decision points of scale up or scale out during server virtualization. In terms of servers, scaling up is usually favoring a lesser number of more capable systems where scaling out is to add a higher number of relatively less powerful systems.

That argument seems so 2008 now. Today’s infrastructure teams have additional options for scale-out architecture like we’ve not had before. There are a few products that recently I have come across that give additional scale-out options, and resurface this argument. Here are a couple of thoughts about new storage and blade products that can make it quite easy to scale out.

Single large server

The HP ProLiant DL980 G7 system was announced last week. This monster will offer up to 8 processors with up to 8 processor cores and a maximum RAM configuration of 2 TB. I have used ProLiant servers for quite a while and have found that the DL380 is a versatile platform, and have favored this model over the four-processor series. My practice has evolved to favor this platform for both physical operating system installs as well as virtualization installations. Use cases for this new eight-processor system include enterprise data warehouses, virtualization host hardware, or anything else that needs an incredible amount of compute power locally. Figure A below shows the HP ProLiant DL980 G7 server: Figure A Figure A

Versatility with blades

The use of blade servers can be an entirely separate debate by itself. Many administrators find benefits in blades when advanced management features and backplane benefits for intra-blade communication are considered. Blades are providing more features today, however. Last week at HP’s Technology Forum, the P4800 BladeSystem was showcased with a SAN and blade host infrastructure. This is targeted at virtual desktop infrastructure installations that scale up to hold the entire infrastructure in a logical bundle that includes hosts, storage, and interconnect networking.

Even more versatility with blades

Recently, I attended the inaugural Hitachi Data Systems (HDS) Geek Day. HDS Geek Day was a social media event for bloggers and HDS engineers. There was a session dedicated to scaling out versus scaling up and HDS blade computing products. We learned that HDS will soon market their blade products in the U.S.

The BladeSymphony series may, at face value, appear like any other blade chassis of eight blades in a 10U enclosure. Blade systems are always feature-rich compared to their chassis server counterparts, but this series has one important distinction. The BladeSymphony series supports logical portioning (LPAR) and hardware virtualization referred to as Virtage, which allow administrators the choice of assigning eight blades individually or assigning all of the resources as one logical server to make the entire collection of blade resources function as one logical server with 64 cores. This gives the BladeSymphony a unique position in being able to both scale up and scale out. Couple this with available storage blades, and virtually every workload seems possible. Figure B shows the BladeSymphony 2000 system: Figure B

Figure B

Unlike their competition, Hitachi can pull a trick out of their hat due to the fact that they have direct access to the supply chain of many electronic components. In fact, HDS Senior Director of Product Strategy Michael Hay said during the event that their direct access (as the OEM) to underlying component offload technologies can be the differentiator in comparing commodity components amongst different vendors.

Scale up or scale out?

These products are just what I have come across in the last few weeks that make me circle back to this age-old decision. Where are you on scaling up or scaling out? Consider also entry cost to scaling out. That has historically been my biggest obstacle to scaling out, as the unit cost of scaling up becomes easier to absorb. Please share your comments on scaling out or up below.

About

Rick Vanover is a software strategy specialist for Veeam Software, based in Columbus, Ohio. Rick has years of IT experience and focuses on virtualization, Windows-based server administration, and system hardware.

3 comments
ydecelles
ydecelles

I guess it depends on what you are scaling and with which supplier ... take virtualization, if you scale up, with HP, considering their price for 16 GB DIMM vs 8 GB DIMM and the pricing for larger blades and extra modules for better I/O channels you might end-up with a big price differentiator between scale-out and scale-up unless you can really increase density. Then again, with new technology such as IBM Ex5 that may not be such an issues... There's also another issue to consider, how risk tolerant are you. Scaling-up usually means higher impact in case of failure (unless you want to spend lots of $$$ on redundancy). So I guess, as I said the answer is 'it depends'. Depends on how risk adverse/tolerant your company is, how much you are vendor locked already, etc...

matthew
matthew

My enterprise is a VMware shop. We started with blades (2 sockets 2 cores 8 GB RAM) six years ago and tried to scale out with them. Next we scaled up with IBM 8 socket rack mount servers - 2 four socket server interconnected (dual cores, 64 and later 128 GB RAM when the OS supported it). Now we are scaling out with blades (4 sockets 12 cores 192 GB RAM) to keep the failure domain less than 200 VMs. We have faced this decision every hardware refresh cycle, the clear answer for us has been re-evaluate every time and since we have the fiber channel SAN infrastructure we are able to go either way.

howlingengines
howlingengines

These other factors mean that you can't just go with 'one of each'.So a 'scaled-up' box will need a partner(maybe more-test/staging etc).In steady-state it's still advisable to have this partner running something even if it's just being 'trickle-fed' status and config info.So you are already slightly scaled out.As usual , a common ground - scaled out network/security/web servers , scaled up app and data servers.The Hitachi picture does reduce this to functional rather than hardware decisions.