The same has also been said in this SliceHost vs. Linode comparison — x86_64 simply uses more memory than the plain old x86. So now you not only paying the same for less memory ($20/month gets you 256MB on SliceHost vs. 384MB on Linode), your applications are also using more memory due to its 64 bit architecture — enough to force you to step up the plan. One of my friends has his WordPress website running at SliceHost. It’s a typical LAMB setup with no control panel, but originally I thought a 256MB Slice would suffice. Apparently not, and the amount of swapping due to possibly bad-optimisation and fat 30+MB Apache processes force him to upgrade to a 512MB slice. But 512MB just to run a low-medium WordPress site? That’s pathetic.

I too have similar experience. Two Ubuntu boxes. One 32 bit at VPSLink and one 64 bit at SliceHost. Both running pretty much my standard LAMP stack serving WordPress and Drupal sites (except Nginx instead of Apache). No opcode cache loaded. Minimum number of extensions. A php-cgi process is around 35MB VSZ and 14MB RSS on 32 bit, but 120MB VSZ and 25MB RSS on 64 bit. That means I can run almost twice the number of FastCGI processes, which can be beneficial on a busy site.

I guess for a standard web app stack, unless I really have a specific need for 64 bit, I’ll probably stick to that plain old 32 bit.

So basically I asked a few questions in the forums (I am “scotty” on WHT.au) –

• Q: Virtualisation? A: Combination of MSFT and Xen
• Q: Debian or Ubuntu? A: Yes
• Q: Data Centre? A: Sydney (later I found out it’s Equinix)
• Q: You insane? A: (1) Limited offer (2) Bandwidth is subsidised (I guess subsidised == oversold)

Well. It’s only around 40 bucks so a few days later I signed up to give it a test. Almost 2 months later I am now writing this review.

VPS Plans — Cheap Australian Bandwidth

Rule of thumb #1 — when someone said it’s “limited offer”, never believe in it! At least I don’t mind in this case, as their “special price” is still available today, 2^nd month into our new financial year. Here’s their pricing table (taken from their website):

$40/month for 100GB is hard to beat, and $70/month for 500GB (AUD$0.14/GB) is way below cost. Storage is on RAID, and the amount of memory is more than enough for what I intended to use this VPS for — serving static data to Australian visitors (as an alternative to using a CDN like CacheFly). I planned to just install Nginx and let it serve Javascript, CSS and small images.

Signing Up

Signing up was straight forward although a bit of manual process is involved. Within 10 minutes of paying by credit card, Nathan from NCL emailed me asking which OS I would like to install. “Ubuntu 8.10″ I replied, and then got the welcome email with IP address and “admin” password within two hours — on a Sunday evening at 10pm!

Although the account activation is not instant, NCL and Nathan still got my thumb up for the fast process.

Initial Impression

So I have now received my IP address and the password, what am I waiting for?!

There is no SSH access!

There is nothing listening on port 22. Having used several unmanaged Linux VPS for the last two and half years and it’s the first time that I was given a freshly-installed Linux VPS without SSH access. It turns out that instead of installing the minimum Ubuntu Server, the standard Ubuntu was installed. On top of that they have also installed Webmin + other packages that I don’t intend to use (Apache, mod_php, etc). The password in the welcome email was actually the Webmin password.

So for the next hour I was trying to figure out how to log into Webmin and then install the OpenSSH server. A bit of fiddling around I have finally got the Bash prompt as root, and the first thing I did was apt-get remove all the extra packages to strip the VPS back to bare minimum.

Then I tried to check out what kind of hardware I got.

$ uname -a
Linux localhost 2.6.24-16-server #1 SMP Thu Apr 10 13:58:00 UTC 2008 i686 GNU/Linux
$ cat /proc/cpuinfo
processor       : 0
vendor_id       : GenuineIntel
cpu family      : 15
model           : 4
model name      : Intel(R) Xeon(TM) CPU 3.40GHz
stepping        : 8
cpu MHz         : 3400.777
cache size      : 1024 KB
fdiv_bug        : no
hlt_bug         : no
f00f_bug        : no
coma_bug        : no
fpu             : yes
fpu_exception   : yes
cpuid level     : 5
wp              : yes
flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss constant_tsc up pebs bts sync_rdtsc pni
bogomips        : 6857.87
clflush size    : 64

Well. Looks like the previous era of Xeon processor, and I only have access to one of the cores. It is running 32bit Linux with kernel 2.6.24. Here is the result of disk performance test:

# hdparm -tT /dev/sda1

/dev/sda1:
 Timing cached reads:   782 MB in  2.00 seconds = 391.53 MB/sec
 Timing buffered disk reads:   70 MB in  3.07 seconds =  22.81 MB/seca

# ./seeker /dev/sda1
Seeker v2.0, 2007-01-15, http://www.linuxinsight.com/how_fast_is_your_disk.html
Benchmarking /dev/sda1 [19751MB], wait 30 seconds..............................
Results: 501 seeks/second, 1.99 ms random access time

Interesting result! The throughput is a little bit disappointing for a RAID storage but 2ms seek time on a VPS?! You have to do a RAID10 on 15k RPM drives or SSD to get that! Or maybe the virtualisation layer is doing some tricks on me?!

Not Xen, Not MSFT — It’s a VMWare

It turns out, that my VPS is not on a Xen node. Nor on a MS Virtual Server/Hyper-V node. In fact it is running VMWare.

# lspci
00:00.0 Host bridge: Intel Corporation 440BX/ZX/DX - 82443BX/ZX/DX Host bridge (rev 01)
00:01.0 PCI bridge: Intel Corporation 440BX/ZX/DX - 82443BX/ZX/DX AGP bridge (rev 01)
00:07.0 ISA bridge: Intel Corporation 82371AB/EB/MB PIIX4 ISA (rev 08)
00:07.1 IDE interface: Intel Corporation 82371AB/EB/MB PIIX4 IDE (rev 01)
00:07.3 Bridge: Intel Corporation 82371AB/EB/MB PIIX4 ACPI (rev 08)
00:0f.0 VGA compatible controller: VMware Inc Abstract SVGA II Adapter
00:10.0 SCSI storage controller: LSI Logic / Symbios Logic 53c1030 PCI-X Fusion-MPT Dual Ultra320 SCSI (rev 01)
00:11.0 PCI bridge: VMware Inc Unknown device 0790 (rev 02)
02:00.0 Ethernet controller: Advanced Micro Devices [AMD] 79c970 [PCnet32 LANCE] (rev 10)

When you list out all the processes, you can also see vmware-guestd running. Is NCL Hosting running ESXi? Infrastructure? I have no idea but at least the VPS performs alright over the last 2 months.

I think the main issue I have with VMWare is the leaky vmware-guestd. It uses quite a bit of CPU, and the memory usage grows from 1.2MB to 20MB within a month. I have to run /etc/init.d/vmware-tools restart to claim the memory back every now and then although I have no idea whether it will introduce any ill effect on a VMWare-based VPS.

2 Months Later…

So how is it after 2 months? I installed Nginx to serve up static data so I am not running anything CPU intensive. The load pretty much stays at 0 most of the time. The VPS itself is rock stable — no reboots, no downtime, the network is always up. Basically I got nothing to complain about.

As I am only serving the Javascript, CSS and small images, I used around 30GB last month so I have not yet had a chance to test the limit. Bandwidth wise it seems to be buying from the Wholesale Communication Group, and I am getting Teleglobe.net for international traffic and Optus for Australia-wide traffic. Bandwidth also seems to be capped at 10Mbps which is fine for me. From a work server I had at Optus DC at Ultimo, I can fully saturated the 10Mbps link. However from both my home and work ADSL2+ link I can’t even sustain 1Mbps, for some reason that I have not yet investigated.

Overall it is pretty much “set and forget” and the server has no problem handling the traffic. However I am still hesitated to bring my site back to Australia (after moving it to US earlier this year), as I think I still trust Linode more if a disaster strikes.

However if you have high bandwidth requirement in Australia, I doubt you can find anyone cheaper.

Before I go on, I feel that I might need to make some disclaimer. VPSLink has also been a regular sponsor of this blog since June this year (in case you have not spotted their skyscraper ads on the right), but I will try to keep my review unbiased :) I was going to give their Xen beta program a try any way when it was announced 2 months ago, but was too busy to do so.

The Product

VPSLink launched their Xen VPS hosting product a month ago after one month of public beta testing. You can read about their Xen Virtual Private Server here:

New Features available in XEN:

• Swap space
• Full control of all iptables modules
• Loadable kernel modules (please note you can not run a fully custom kernel)
• Access to remote console for troubleshooting

XEN differs greatly than OpenVZ and there are several factors to keep in mind when selecting a XEN based VPS account.

Then it listed out a few things to consider to choose between Xen and OpenVZ. None of those matters to me anyway, and it failed to explain how Xen and OpenVZ are in fact intrinsically different, especially how memory model differences will turn a perfectly fine application on Xen to instability on OpenVZ, which I sort-of explained here.

The plan price is nowhere near low-end, however they are very cheap consider VPSLink/Spry’s reputation. My test VPS is a “Link-3″, which comes with:

• 10GB disk space
• 300GB bandwidth
• 1 dedicated IP
• 256MB dedicated RAM

It costs between $20.79 – $24.95, depending on the length of your billing cycle.

Signing Up and Setting Up

Signing up is the easy bit — it has all been done for me when I asked Cameron for a test VPS! However from my previous sign up experience (over 17 months ago), the provisioning is instantaneous (if you pass through their automated fraud check). You should be receiving instructions within minutes of payment.

VPSLink uses its internally developed control panel for both OpenVZ and Xen VPS. It is easy to navigate and the layout is pretty much task-driven. The first things I do were logging into control panel, select my new VPS, click on “Manage OS” and then “Install OS”. Ubuntu 7.10 came out around 2 weeks ago, and it was already available on the list of OS images to install. “Why not give the latest Ubuntu a try as well?” so I clicked on Submit and waited.

It took only a few minutes for VPSLink to rebuild the server with a base Ubuntu installation (which is one thing I really like about VPS). An email is sent out containing the root password, and I am in with a root shell in minutes!

Software and Hardware

As it is paravirtualisation and you cannot run your own kernel, VPSLink dictates the kernel version on all the Xen VPS. Here is the uname -a output.

# uname -a
Linux vpslink 2.6.18-8.1.15.el5xen #1 SMP Mon Oct 22 09:33:52 EDT 2007 i686 GNU/Linux

Looks like it is running the kernel from CentOS 5 (or RHEL 5). You can build your own kernel modules but if you are not running CentOS 5, you’ll need to grab the Linux 2.6.18 kernel source and compile your kernel module against that.

Hardware wise it is not that fancy. Basically VPSLink has two classes of servers — dual core single proc for Link-1 to Link-3 customers, and dual core dual proc Xen for Link-4 to Link-6 customers. So what do I get with my Link-3 account?

# cat /proc/cpuinfo
processor       : 0
vendor_id       : GenuineIntel
cpu family      : 6
model           : 15
model name      : Intel(R) Core(TM)2 CPU          4400  @ 2.00GHz
stepping        : 2
cpu MHz         : 1994.998
cache size      : 2048 KB
fdiv_bug        : no
hlt_bug         : no
f00f_bug        : no
coma_bug        : no
fpu             : yes
fpu_exception   : yes
cpuid level     : 10
wp              : yes
flags           : fpu tsc msr pae mce cx8 apic mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe nx lm constant_tsc up pni monitor ds_cpl est tm2 cx16 xtpr lahf_lm
bogomips        : 4989.56

Basically an Intel Core 2 Duo E4400 CPU with 2MB L2 cache running at 2.0Ghz, i.e. “low-end”. Moreover, my VPS has only been assigned to one of the two CPU cores so it is impossible to burst my CPU usage to the full potential of the CPU. A single E4400 core is still plenty of power for a 256MB VPS and my neighbours will be happy to know that my “make -j4″ will not affect the whole node.

Still, while a single E4400 core packs enough power for a low-end VPS (and VPSLink probably keeps the number of domU low), it is nothing to be boasted about. I have no CPU performance problem though over these two weeks, as web-serving VPS usually have their bottleneck in disk IO.

There are plenty of disk IO bench mark programs, but I just use something that is readily available — hdparm.

# hdparm --direct -t /dev/sda1

/dev/sda1:
Timing O_DIRECT disk reads:  486 MB in  3.00 seconds = 161.88 MB/sec

# hdparm -T /dev/sda1

/dev/sda1:
Timing cached reads:   2330 MB in  2.00 seconds = 1165.09 MB/sec

Direct disk read at 161 MB/sec — that’s pretty impressive! Reading from disk cache at 1.1GB/sec — Jeez! It probably has the best disk read speed comparing with my other VPS.

Network

VPSLink is in Seattle, inside Spry’s own data centre. Network performance has been great around the clock, and there has not been any downtime that has been detected (polling at 5 minutes interval).

But you know what is good about hosting in Seattle? Three words — Amazon Web Services. Amazon is also in Seattle, and if you are a heavy user of services like Amazon S3, or want a static IP front-end for your fleet of Amazon EC2 slaves, the latency and bandwidth between your servers and Amazon can make a huge difference.

Check out these benchmarks.

# tcptraceroute s3.amazonaws.com
Selected device eth0, address 209.40.199.168, port 39897 for outgoing packets
Tracing the path to s3.amazonaws.com (207.171.181.225) on TCP port 80 (www), 30 hops max
 1  64.79.219.1  0.671 ms  0.370 ms  0.433 ms
 2  64.79.223.1  0.429 ms  0.458 ms  0.470 ms
 3  ge1-4.cr01.sea02.mzima.net (72.37.232.33)  9.481 ms  0.933 ms  0.466 ms
 4  xe0-1.cr01.sea01.mzima.net (216.193.255.193)  1.470 ms  10.979 ms  0.912 ms
 5  ge-6-2.car3.Seattle1.Level3.net (4.71.152.1)  0.931 ms  0.943 ms  0.996 ms
 6  ae-12-55.car2.Seattle1.Level3.net (4.68.105.131)  0.951 ms  0.965 ms  0.976 ms
 7  * * *
 8  185-33.amazon.com (207.171.185.33)  1.459 ms  0.953 ms  0.973 ms
 9  177-159.amazon.com (207.171.177.159)  0.976 ms  1.449 ms  0.968 ms
10  207-171-181-225.amazon.com (207.171.181.225) [open]  0.982 ms  0.946 ms  0.960 ms

# wget http://s3.amazonaws.com/ec2/assets/aws_console_screencast_1.mov
--06:45:35--  http://s3.amazonaws.com/ec2/assets/aws_console_screencast_1.mov
           => `aws_console_screencast_1.mov'
Resolving s3.amazonaws.com... 207.171.183.113
Connecting to s3.amazonaws.com|207.171.183.113|:80... connected.
HTTP request sent, awaiting response... 200 OK
Length: 11,004,220 (10M) [video/quicktime]

100%[==========================================================================>] 11,004,220    11.36M/s

06:45:36 (11.34 MB/s) - `aws_console_screencast_1.mov' saved [11004220/11004220]

Yes, read that! Less than 1ms latency to Amazon S3, and pulls a 10MB movie file down at 11.36MB/sec! Awesome.

Still with Bugs

Great performance aside, somehow I feel that VPSLink’s Xen VPS is still not 100% ready. There has been a few issues with their console panel and OS images.

• Somehow /etc/fstab wasn’t set up properly on the Ubuntu 7.10 image. It is emptied out, so commands like df and mount do not show the root directory is mounted, thus you can’t find out how much space has been used. It wasn’t an issue with Debian 4 OS image that I’ve tried.

• Getty wasn’t initialised on console for my Ubuntu 7.10 image either, but then console access from control panel does not work for me either (always got connection timeout). The screen cast with CentOS shows it’s working fine so it might just be Ubuntu OS image again.
• Other glitches like hostname is not persistent.

None of that are show stoppers, and VPSLink is actively working on resolving these issues. Hopefully they will be resolved soon.

Conclusion

Xen VPS at VPSLink provides solid performance and great network connectivity especially to Amazon Web Services. The low-end VPS might not get state of the art hardware, but disk IO is certainly not the issue in my own experiment.

Moreover, it has a great community and good support. Recommended if you are looking for a West coast VPS.

Xen and OpenVZ — What’s The Difference?

While both Xen and OpenVZ are open source server virtualisation technology, there exists some big differences between the two. I think potential VPS customers might need to check the applications that need to be hosted to determine which one is the preferable virtualisation technology.

On one hand you have Xen, a para-virtualisation platform that gives you much of the dedicated server behaviour. You run your own instance of Linux kernel, you can load your own kernel modules, you have properly virtualised memory, IO and scheduler, and it’s stable and predictable. On the other hand you have OpenVZ, an operating-system level virtualisation system that is just a thin layer on top of the underlying OS. It is simple to understand, has lower overhead, which usually translates to better performance.

VPSLink is offering both OpenVZ and Xen VPS of similar specs at the same price. For example, I have a Link-3 OpenVZ VPS running CentOS 4.5, and it has 256MB memory, 10GB storage space and 300GB data transfer per month. VPSLink has also provided me a Link-3 Xen VPS for my review running the latest Ubuntu 7.10 ,and it too has 256MB memory, 10GB storage and 300GB data transfer per month. Same price for both — $24.95/month if you pay monthly. Now, which one should I buy?

OpenVZ Memory Model

First of all, when VPSLink’s OpenVZ Link-3 says “256MB guaranteed”, it actually means around 232MB of “privvmpages”, 14MB of “kmemsize” and other miscellaneous resources. When an application calls malloc(), the allocated amount will be added to “privvmpages”. However when “privvmpages” hits the limit, malloc() will fail with a NULL. When the host server ran out of memory, then processes in VE (virtual environment, OpenVZ’s term for a VPS) that exceeded “oomguarpages” will be terminated, although I do not think it will ever apply to VPSLink.

There are a few problems and a few advantages with OpenVZ’s approach to memory management. One of the biggest problem is the amount of memory an application uses and the amount of memory an application allocates is actually different, and the difference can vary a lot depending on the application. Take Java for example, it usually allocates a huge chunk of memory — usually everything it can see the host node has — but it might only use/commit a small fraction of allocated memory. It can usually render a Java program unusable as you will pretty much hit the privvmpages limit straight away. A bit of tweaking on bootstrapping parameters might fix it, but it is definitely not as clean as Xen or a dedicated server. In fact, almost all applications that use internal memory allocator suffer from this issue under OpenVZ.

Then there are issues associated with /proc/meminfo itself. While OpenVZ has already provided a way to virtualise it, “free” command on my OpenVZ VPS at VPSLink still shows the memory size of the host node. It makes some tasks, like running Java or heavy compilation with gcc almost impossible on a small VPS.

The advantage of OpenVZ’s memory model is that it is simple to understand — you are pretty much limited by only privvmpages on a VPSLink OpenVZ VPS. Unlike a dedicated server or Xen, your disk cache and your buffered pages are not counted against your overall memory usage. Therefore on an under-sold OpenVZ system with lots of cache and buffer memory on the host server, it might actually perform better than a similar spec’ed Xen VPS.

Xen Memory Model

Memory model for Xen VPS is much easier to explain. A 256MB Xen VPS is just like a 256MB dedicated server — that segment of memory is reserved for this VPS only, and no other VPS nodes can touch it. Like a real dedicated server, it counts only resident pages, i.e. only the memory blocks are allocated and used.

Moreover, what happens when you run out of memory? You VPS starts to swap. Each VPSLink Xen VPS comes with a swap partition that is twice the size of memory. When your application requires more memory, least-often used pages will be swapped out to make more rooms. Therefore a 256MB Xen VPS actually has 768MB of total memory (256MB RAM + 512MB swap), and believe me, swap space is very useful to handle that sudden spike of demand.

So Xen is always much better than OpenVZ? Not quite. While your 256MB VPS can theoretical use up to 768MB of memory, in reality

1. Kernel, cache, buffer — they all take up memory.
2. Swapping kills performance.

Yes, you can tune the swappiness so you can keep on reducing cached and buffers without touching swap, however performance will suck. On the other hand you can bump up memory usage on an OpenVZ VPS all the way to the privvmpages limit without much degrading of performance, provided the host node still has room to spare. It is a good thing but can also a bad thing, which I will explain later.

Performance vs. Predictability

At the end it comes down to performance and predictability, and my preference has always been with Xen.

While Xen has more overhead thus possibly slower VPS, its out-of-memory behaviour is much more predictable than OpenVZ, and this predictability won me over. As I have already said, that OpenVZ will continue to perform well when its memory usage approaches the limit. However, if privvmpages has been exhausted, the next malloc() will return a NULL pointer, and depending on how the applications handle NULL pointer, they either die gracefully or die with a segmentation fault (usually the later). It is like driving at 100km/hour and then suddenly hits a brick wall. You do not even know that there is a memory shortage issue because everything just sails so smoothly.

On the other hand, when a Xen VPS used up free memory, it will start taking memory from buffers and cached pages. Then it will start swapping. And then it will finally die when the last bit of swap partition gets exhausted. Performance will start to degrade when it started swapping. The load will go up, and the server will get less and less responsive. Your Xen VPS will spend more of its time swapping pages in and out than actually handling tasks. Even if it dies at the end, it will be a very noticeable long struggle than a head-on smash…

… just like a dedicated server!

My preference? Predictable performance. I’ll rather have my sites slowing down to its knees, than having it crash and burn when the memory is exhausted.

What About Burstable Memory in OpenVZ?

“Burstable memory” in OpenVZ is overrated IMHO, as it makes the behaviour of your VPS even less predictable. It is often advised to set privvmpages (burstable amount) at twice the amount as vmguarpages (guaranteed amount) as allocated amount vs. used amount is usually 2:1. However it is not always the case. At work where we had lots of Java development it’s a bit like 5:1, but on my VPSLink OpenVZ VPS where it is mostly Lighttpd, MySQL and PHP, the ratio is about as low as 1.45:1.

Therefore out-of-memory could still happen when you have burstable (privvmpages) set to twice the guaranteed (vmguarpages). On the other hand, VPSLink’s “no burst, no swap” policy, i.e. making burstable amount the same as guaranteed amount, actually gives each VPS less memory to play with, at least it guarantees that when OOM occurs, no VE will be held responsible as all of them will be under their guaranteed limit.

Conclusion

By looking at OpenVZ and Xen’s memory model and their handling of out of memory situation, I am leaning towards Xen, if I am going out to buy a VPS now. Especially when they are sold at the same price per month at VPSLink, I personally cannot see why someone would prefer OpenVZ over Xen.

Except if there is a discount on OpenVZ VPS :)

I have seen a few promotions coupons from VPSLink (last one being 31% discount from Halloween) but none of them can be used on VPSLink’s Xen VPS. I for one will switch from OpenVZ to their Xen VPS in a heart beat, except I am still on the initial 50% discount for my OpenVZ VPS that is just too hard to let go :)

Alright. Enough Xen vs. OpenVZ. I will be reviewing VPSLink’s new Xen VPS hosting in my next post, on what’s good (+ what’s not so good) about them.

I mean it’s pretty obvious why some would want to buy VPS; for the value of course compared to true dedicated server offerings. It’s just a perfect next step from shared without having to incur all the expense of a dedicated. (Emphasis mine)

Is virtual private server the “perfect next step from shared hosting”? Is VPS the logical step-stone between $10/month shared hostings and $100+/month dedicated servers? Is it something you buy when your hosting company kicks you out for resource overage, and are too poor to buy your own “real” server? What do you think?

One thing I found interesting is where is VPS marketing at. If you are a hosting company, and are selling both shared hosting and dedicated servers, what will you put on their specification pages?

It is very obvious that these two products are for two completely different markets. You don’t see a lot of shared hosting company boasting about their RAID10 disk arrays and quad-core Xeons, nor many dedicated server companies see “unlimited domains” and “blog software included” as necessary in marketing their products.

Now, go down to the VPS offering section of WebHostingTalk, and check out how VPS providers market their products. Now tell me which side of the table above resembles the VPS specifications?

From my own understanding, VPS is more of a “small dedicated server”, than a “big shared hosting account”. It is intended for those on dedicated servers to scale down (server consolidation, for example), than for those on crowded shared hosting to scale up. It has all the characteristics of a dedicated server — more so if the provider is using hardware/para-virtualization. It also requires similar commitment as a dedicated server. Application installation, security, administration, system tune up/optimization, etc.

Personally I find it is a perfect hosting solution for me, who want the power of root, but don’t want to pay for (nor require) the whole box.

However, I have seen VPS frequently recommended as the next level up from shared hosting to novice who happens to have a busy website. Because of the required server administration skills, people either come back 2 days later complaining how crappy VPS is, or they ended up purchasing cPanel + 256Mb extra memory + managed service in order to make that VPS “functional”. I think that is just so wrong.

No, I don’t think VPS is a suitable candidate for hosting resellers. VPS is one level down from dedicated servers, but one level up from shared hosting should be managed shared hosting with cluster servers instead (MediaTemple’s Grid Server and Rackspace’s Mosso for example). You don’t need to be a shell whiz-bang to use them, and the cluster servers, if implemented correctly, should easily scale up when the traffic soars.

1. Acknowledge MySQL issues in the current grid server platform.
2. Talk about their old load balancing system for MySQL, and why it did not work.
3. Discuss their new VPS based MySQL solution that will be switched on in March.

Scaling databases on a shared multi-tenant system has to be one of the most challenging exercise for system administrators, and I am sure the Media Temple would agree. Because of the way web applications are usually written these days, scaling your PHP apps can be trivial. PHP’s share nothing architecture makes scalability native to the language itself, and you can simply add more boxes to take care the load on the web servers.

However, with “share-nothing”, the design is that all states are pushed into the database, and the assumption is that database is scalable. Is the database scalable? That alone is worth a separate discussion article. To cut the story short, you cannot simply add more hardware to scale up a busy database server. It usually requires the application to optimise the way data are queried. Setting up replication. Distribute heavy queries. Partition the dataset. Rewrite big queries into smaller ones in case of MySQL. Etc etc.

With shared hosting, you’ll get even more problems. As a web hosting provider you need to optimise your database to cater for your customers who might install random or home-made apps with badly designed queries.

Media Temple’s old design was pretty interesting — a load balancer for MySQL database clusters. From the description, it looks like it constantly monitor all MySQL servers in the system, and live-migrate busy databases onto servers that are less loaded. However it does not work as hostile DB users can still bring the entire database server to its knees, where anyone else on the same server will suffer.

Media Temple’s new approach is even more interesting — they will put MySQL into “containers”, i.e. a VPS-like environment with restricted CPU and RAM. From the description I think the new MySQL container approach will be even easier to manage than the current load balancing cluster approach.

Each account would have its own MySQL server running inside a VPS. Instead of continuously monitoring the load and move the offending DBs around, they can just rely on the virtualisation software to keep each instance in check. A busy user will only crash his own MySQL container and won’t affect other accounts. Likewise, a light user should always receive guaranteed CPU and IO.

That’s quite revolutionary for shared hosts. Well done! It also demonstrated a great use case for VPS — partition otherwise shared resources.

I will go for a VPS, if I have limited budget, and have a busy site to host.

For example, you happen to have $70/month to get something semi-decent. On one hand, you have a cheap dedicated server from a ServerPronto (one of the cheapest dedicated server providers), on the other hand, you have a top-level Xen setup on SliceHost.

They look “somehow” comparable, aren’t they? SliceHost has slightly better CPU power — remember their physical box has quad core but they are shared with 4-7 other 1024slices on those 8Gb Opteron boxes, plus very minimal overhead on Xen. Twice the memory will be very helpful to keep your Apache processes from swapping in and out of disks. Storage and transfer are lacking, but (1) they are soft-limit on SliceHost, and (2) they don’t oversell bandwidth.

Note that I have no experience with ServerPronto. I did have some recent experience with SliceHost. However if I do have a budget of $70/mo I would be giving ServerAxis a try, who is actually cheaper with their better-equipped VPSs.

Alright — so they are on-par on paper. But let me show you why getting a VPS is much better.

• CPU is burstable. When all the virtual servers are fighting for CPU, you might feel like running on the same speed as a AMD 2600+. However, when there are idle cycles, you are free to grab them. So the end result can be much faster than that 3-year-old chip in that cheap dedicated, sometimes.

• I/O is burstable. Slow disk IO is usually the most criticised short-coming of VPS, but I noticed that it usually only happen on crowded Xen or oversold Virtuozzo servers. Well in SliceHost’s case, you don’t get a dedicated 40Gb hard disk for your slice, but on a shared RAID1 storage of much bigger and faster drives. End result? Much faster disk I/O when other VPS’s are not fighting for it. Many other VPS providers use RAID5, RAID10 or 0+1 on very fast SATA, SCSI or SAN volumes.

• Redundant storage. Most VPS have RAID on their drives, because everyone knows — hard drives will die. After quite a few incidents at home, I will not trust my data on anything that is not RAIDed. Not only do they provide redundancy, but also faster recovery if one of the drives does die, to minimise down time.

• Server grade hardware. What do you think your AMD 2600+ box is? A well designed rack unit with redundant power supply and sufficient cooling? Server grade motherboard with ECC memory? Most likely a beige desktop box. Whereas other dual Xeon/Opteron boxes used to host VPSs are definitely better designed and equipped.

• Extra management. Both VPS and dedicated servers are “unmanaged”, but I’ll argue that VPS are more managed than a dedicated server. The host’s administrator will look out for hardware issues. They’ll help you to make sure your VPS boots. Some of them even provide virtual console so you can diagnose issues yourself. Remote console, reboot support, etc can all cost extra on dedicated servers.

• Better for the environment. Not sure whether you care, but you can squeeze in more VPS into the same rack than cheap dedicated servers. Thus less electricity is used. Somehow it is better for the environment, making your providers less depending on the more and more expensive energy price.

Well. I am sure there are more reasons, but all these should have concluded that there’s few reason to get “cheap dedicated” these days. Moreover, the real sweet spot of VPS is less than $50/month, which is enough to get a medium-level VPS to host multiple website or a “single purpose” web application. What kind of dedicated do you expect to get with that amount of money?

Hosts can also take advantage of virtualization. Besides “better for environment”, better CPU utilization, VPS also has advantage of faster provisioning. Most dedicated servers take a few hours to provision a new server request, and longer if hardware is not available. Most VPS provisioning can be scripted and automated, provided there’s enough room on the host node.

Yes, dedicated servers do get more consistent performance. More consistent CPU response. More consistent IO throughput. But, choosing between being consistent with a AMD 2600+ performance, or burstable to dual Opteron 265 performance — I think I know which one I’ll go.

But it is a dwarf in comparison to Amazon’s Elastic Compute Cloud (EC2), which is currently at limited beta.

What is EC2? From Amazon’s page:

Amazon EC2 presents a true virtual computing environment, allowing you to use web service interfaces to requisition machines for use, load them with your custom application environment, manage your network’s access permissions, and run your image using as many or few systems as you desire.

And how is it different from other hosting service? From their FAQ page:

Traditional hosting services generally provide a pre-configured resource for a fixed amount of time and at a predetermined cost. Amazon EC2 differs fundamentally in the flexibility, control and significant cost savings it offers developers, allowing them to treat Amazon EC2 as their own personal data center with the benefit of Amazon.com’s robust infrastructure.

Basically, it is a virtual private server hosting service that (1) servers can be instantiated and destroyed in real time with web service API calls (2) you only pay for the resource you are utilising. Pricing structure is pretty easy to understand:

1. $0.10 per hour when your VPS is running.
2. $0.20 per GB of data transferred outside of Amazon.
3. $0.15 per GB per month of data stored.

Each VPS instantiated has a spec of 1.7Ghz Xeon equivalent, 1.75Gb of RAM, 160Gb of local disk and 250Mb/s network bandwidth, so it is not the $20/month VPS that people rent to run their blogs. So to have a server running 24×7, a 30 day month would cost $72 — and that is not including data transfer and storage. Even with 40Gb of storage and 250Gb of data transfer it would still cost only $128.

However we need to remember that you are only paying what you are using. That means Amazon is not overselling — they don’t promise their customers 1,000Gb of data transfer, hoping that some are only using 20Gb to offset other hungry customers. It is certainly very refreshing from the current dedicated/shared hosting market.

Also from the pricing structure above, (2) and (3) come directly from Amazon S3, a storage infrastructure that is scalable, reliable, fast and cheap. Like another SAN-backed VPS hosting, if the physical server that hosts your VPS happen to burn in flames, it is easy to re-instantiate the same image on another physical server to reduce down time. It is not the luxury you can get from other $100/month dedicated servers.

From its developers’ documentation, EC2 appears to be powered by Xen, but you cannot yet run your own kernel. That seems to contradict the claim that Xen is not yet ready for production use.

What differentiates EC2 from other Xen providers is, this web service API that allows real time provisioning. Need a 20 Xeon server farm for your secret project in the next 10 minutes? Amazon says no problem. Only use them for 10 days? No problem either, and you are only paying for the amount of time you are using. EC2′s API is clearly documented, and from the look it could do some amazing things like grouping instances, etc.

It is going to be very useful when you need that computation power suddenly. For example, need to crack that hash getting Slashdotted or Dugged. Just use the API to create a few more instances. Provided that your software can scale, adding new hardware has never been easier.

With S3, EC2 and other services, Amazon has clearly taken the path to be centrepiece to the developers, like how Google is to average web users. As software developer, you just need to focus on your application’s scalability, and Amazon will take care of the scalability of computation and storage.

On the other hand, this service might not yet be suitable for general web hosting — not when it is in beta anyway. Your VPS’s IP is allocated on DHCP so no static IP address, which cannot be easily used to host web and mail services without going down to the dynamic DNS hack. Also there is no persistent storage other than S3 — rebooting your VPS and you’ll be surprised to find out all data you kept on the 160Gb partition has disappeared. Otherwise, I can forsee EC2 as a popular platform to build shared hosting, as you can easily oversell on non-overselling resources.

Too bad beta has already been closed, but I am sure I will be eagerly waiting for this one.