Ten tips on x86 Server Buying
Caveat, Not everyone will agree to these observations. Consider this also as a honest buyer’s guide to server market.Almost all of these Gyan has come from years of consulting, buying, configuring and selling servers and making many mistakes often at the expense of others.
Some Data Crunching
I had access to server purchase records of over 130 Indian customers who have ended up buying close to 800 servers across past four years. These include File Servers, Application servers, Web servers, and even heavily Virtualized boxes. Almost sixty percent of these customers have purchased less than 3 servers during the same period. The maximum number was a customer who has purchased 34 during the same period.
Here are some vital statistics which I gathered from this data.
- Nearly 46 percent of the users never altered the hardware configuration after their purchase.
- Just 8 percent of the users who purchased a dual processor capable server with single processor opted to upgrade the processor to make it run in dual processor mode during the same period.
- Almost 50 percent upgraded their memory usually by adding 4 to 16 GB of memory to the configuration they purchased.
- About 20 percent added newer drives or upgraded drive technologies like shifting from SATA to SAS.
- Less than 15 percent actually had to get their hardware fixed during the first three years (warranty period) of ownership. In most cases faults were with the SMPS or the motherboard. Rest of the cases where customers had a down time was a software related issue.
- Almost 20 percent opted to buy new servers or switch to cloud because they were not happy.
The accuracy of facts e) and f) are may be questionable. The data I have gathered is from a reseller and couple of systems integrators. Their CRM does not capture some of these details accurately. So factor in a variation of at least 10 percent.
Based on these and my own experience I have compiled these 10 tips.
Never buy a Multiprocessor capable server if you don’t need to
Often sales guys end up selling you stuff you don’t need ever. I have met several customers who have ended buying high end servers which are extremely under-utilized. Never buy a dual processor server when all you need is the power of a single processor server.
A Xeon E3-1231v3 processor which is priced at $250 (Rs 16500) which ships in an entry level server priced about Rs 50,000($600) will perform better than a dual processor server sporting a single Intel Xeon E5-2640 v2 clocking 2.4 Ghz which would be seen in a server priced typically at Rs 120,000($1600) while rest of the configurations being similar. Note, I am talking of dual processor configured with a single processor and a single processor server. I would recommend you to add more memory and invest in faster drives and more redundant configuration.
This is fundamentally because single processor Xeons are clocked at higher speeds. 80 percent of applications that you run on a server is not truly multi-thread enabled. So single thread performance is quite critical, and is generally provided very well by single processor Xeons.
If your application is not CPU intensive, and you don’t plan to run more than two virtual machines, and if all you realistically need is 32 GB RAM or lesser do not look beyond a single processor server. With the extra budgets invest in faster and more drives.
Additional Tip:Think SSD. Think RAID 10. We will come to that later.
2) Never buy a high end enterprise server if you don’t need to.
I think some CFOs will start loving me for providing these two tips. Again often I have seen customers buying an HP Proliant DL 380 or Dell PowerEdge R730/830 where all they require is Proliant DL 80 G9 or a Dell PowerEdge R430. This often ends up in savings of Rs 50,000. So in case you don’t plan to host more than 128 GB of memory on a single box or do not plan to have more than four hard drives, then bet on the entry level servers in the dual processor models. The difference between a Dell R730 and Dell R430 is that the more expensive model has more memory slots, expansion slots and more bays. If you don’t visualize the need for these extras today, you are hardly likely to need them in next two years. In that time Intel would have shipped two more generation of processors and several new server technologies would have evolved.
Additional Tip: In three years evolving server technologies will offer you double the performance.
3) Always Virtualize if you can
Caveat. Not all workloads are meant for virtualization. But if there is scope for virtualization, and an opportunity to cut down the number of servers, do that. Most non virtualized servers have a CPU utilization that is in single digits. Remember there is no thumb rule for the number of virtual machines which you can spin off a physical server. It depends on applications and workloads.
Additional Tip: Always ensure that every VM is backed up. Invest either in a good VM commercial backup software, or have someone set up that on top grade open source software.
4) Always compartmentalize servers
Unless you are a small business where economic constraints play a very big role, you need to separate servers. Never run application servers, database servers, file servers and backup servers all on the same physical or virtual machine. Always run them off separate machines.
You never know when one infected file formats the entire hard drive of a server. Talk today with your systems team, and ensure that they do not commit this mistake. If already committed, see that they correct.
Additional Tip: Monitor server performance
5) Never invest in RAID 5.
RAID 5 is one of the most popular of RAID configurations. The industry loves selling RAID 5 and RAID 6 configurations. These servers are often expensive as hardware RAID cards that are essential for a stable RAID 5 implementation is expensive. Customers often get inspired by the tech jargon, and the idea that you can actually get a utilization value at expense of just an extra drive is appealing..
But please remember these golden rules.
- Hard drives are cheap, your data is not.
- Just about 50% of RAID 5 crashes have seen data being recovered. Others had to either write off the data loss, or had to spend 20 times the cost of the hard drives to recover data and wait for days chasing a data recovery specialist.
- RAID 0 is the fastest among RAID 0,1,5,10. But it is also the most unreliable. A drive lost means all data lost with RAID 0.
- RAID 5 is the second most unreliable. RAID 1 is fairly reliable but your performance will be slow.
- RAID 10 is the most reliable and performs the best. If you know how to optimize software RAID 10, your performance will be better than a RAID 5. running on a commodity hardware RAID card
- Remember an HP, Lenovo or Dell will promptly replace the hard drives, but they will give no guarantee on the data. So invest in RAID 10 always. If not RAID 1. No RAID 5.
You don’t trust me. Pay a visit here. Read from the experts.
Additional Tip: Not sure how to optimize RAID 10. Call experts. If on Linux, consult me :)
6) Invest in SSD and in as low capacities as possible if you can only afford spinning drives
Unless you are running a backup, file or archival server always invest in SSD drives. Yes SSD drives are not the most reliable, and their performance does degrade over time. But in most cases they should be good for next two to three years, and after that buy new drives. Throw the old ones, or reformat those and run them on old desktops.
If you are only able to afford traditional drives, please invest in as lower capacities as possible. Remember a 500 GB SATA is likely to run faster than a 1 TB SATA, and a 1 TB faster than a 2 TB. The larger the HDD capacity more time in needs to spin and find out where data is. Pay also an attention on the drive caches.
This is where you need to invest in tools or scripts to clear data and ensure that you can optimize with as low capacity as possible. In most cases simple bookkeeping helps you keep servers clean.
The low capacity rule however is not applicable to SSD drives, where the numbers you need to focus is the number of Read/Write IOPs.
Still worried about reliability of SSD drives. Remember the rules. Always ship with RAID 10, you cannot go wrong.
Additional Tip: Buy drives with 5 year warranty. They will surely last for 3 years.
7) Bet on Racks if physical server count is more than 4. Think Blades if the number of physical servers cross 20.
Now there is some scientific thinking behind all this, though this thumb rule will not work in all cases. And I will not explain how I arrived at these numbers, but the parameters I have considered to take this decision include.
- Cost difference between tower, racks and blades.
- Additional Infrastructure costs and real estate space considerations
- Management software that gets shipped with Blade Chassis and high end racks.
- Running costs including power
I know some circumstances are there where you may want to invest in a blade chassis when the number of physical processors is as low as 16 processors. Again, some exceptions aside bet on my rule.
Additional Tip:Check on upgradeability of Blades. The truth is many that are shipping are simply not upgradeable.
8) Bet on Gold rated SMPS of lower capacities
Power bills account for 60 percent running costs of a server room or a data center. Lower the VA rating of the server the lower your bills are likely to be especially when you are getting the servers co-located.
However many server vendors ship 750/800/900 W SMPS as a standard option with some models. Never buy those if you don’t need them. 145 W is the highest TDP rating for a E5 Xeon processor.Average is 105W.Add a 5 W for each cooler, to host two processors you don’t need more than 220 W. Motherboard pulls in another 95-150 W. Average is 135 W.Memory sticks about 2 W and each spinning drives about 12.5 W each.
Mind you TDP rating does not mean the components pull in the maximum power always. SSD drives pull typically 2 W of power, which means a server with dual processors, four SSD drives on RAID 10 and 128 GB RAM should pull in less than 400 W of power at the maximum. Hence a 450 W SMPS suffices in most cases, when you don’t need to host a heavy duty RAID or HBA card.
Additional Tip: Invest in Redundant Power Supplies if your uptime is very important.
9) Remember Best of Breed is always better than Closed Stacks
Some enterprise customers have a dirty habit. They want a single neck to choke. They prefer a single vendor, because they believe that there is likely to be uniformity. That’s the reason many vendors have launched their own Integrated stacks of server, storage and networking components.
Boy, if you are a bit smarter than Rahul Gandhi, you will know that you are just walking into yet another vendor lock-in. Try not to fall into that trap. Remember any attractive bundled offer always will have some reason behind it.
The best of breed approach means that you always have a choice to keep tinkering with the best available resources. They are often 15 percent cheaper to buy and integrate, and you still have a choice to negotiate a year down the line as you keep growing your infrastructure.
Additional Tip: Read the next tip
10) Buy from the best, not just buy the best
It’s true that server hardware is commodity. Hardware is supported anyhow by the vendor themselves for 3 years. And you can just about source from anywhere. There are a number of stockists out there who is selling these hardware at low margins. Do not buy from them. Buy always from someone who understands technology, who can consult for you. Buy from a true systems integrator.
Remember a true systems integrator will not sell you integrated infrastructure. They will always suggest you best of breed. Only a corporate reseller will suggest you a single brand approach.