Friday, March 21, 2014


NTP-based DDoS attacks are fashionable, currently.

I've coded a little application which quickly scans a network for NTP servers. For those found, it rates them according to their susceptibility to being implicated in an amplification DDoS attack.

Saturday, October 01, 2011

What a dying SFP looks like

Fibre channel (FC) storage is handy, and generally very reliable, in my experience. I certainly do not miss the days of messing around with disks in a server-room. And I like the fact that RAIDs may be cut up into slices (LUNs) which may be shared by many servers, resulting in very efficient use of the disks (if so wanted).

One part about FC that I dislike (in addition to the price tags): SFPs. Why on earth are transceivers not an integral part of a Fibre Channel switch? Having the transceivers be separate units means more electrical contact points, and a potential support mess (it's not hard to imagine a situation where the support contract of an SFP has run out, while the switch itself is still covered).

Anyway: Today, I experienced an defunct SFP, for the first time. The following observations may give a hint of how to discover that an SFP is starting to malfunction. The setup is an IBM DS4800 storage system where port 2 on controller B is connected to port 0 on an IBM TotalStorage SAN32B FC switch (which is an IBM-branded Brocade 5100 switch).

Friday morning at 07:49, in syslog: A few messages like this from the FC switch:
raslogd: 2011/09/30-07:49:07, [SNMP-1008], 2113, WWN 10:... | FID 128, INFO, IBM_2005_B32_B,  The last device change happened at : Fri Sep 30 07:49:01 2011

At the same time the storage system started complaining about "Drive not on preferred path due to ADT/RDAC failover", meaning that at least one server had started using a non-optimal path, most likely due to I/O timeouts on the preferred path. And a first spike in the bad_os count occurred for the FC switch port:

bad_os is a counter which exists in Brocade switches, and possibly others. Brocade describes it as the number of invalid ordered sets received.

At 10:55, in syslog:
raslogd: 2011/09/30-10:55:02, [FW-1424], 2118, WWN 10:... | FID 128, WARNING, IBM_2005_B32_B, Switch status changed from HEALTHY to MARGINAL
At the same time, there was a slightly larger spike in the bad_os graph.
Coinciding: The storage system sent a mail warning about "Data rate negotiation failed" for the port.

At 17:00: The count for bit-traffic flat-lined (graph not shown). I.e.: All traffic had ceased.

At no point did the graphs for C3 discards, encoding errors or CRC errors show any spikes.

The next morning, the involved optical cable was switched; that didn't help. Inserting another SFP helped, leading to the conclusion that the old SFP had started to malfunction.

Morale: Make sure to not just keep spare cables around. A spare SFP should also be kept in stock.
And monitor your systems: A centralized and regularly inspected syslog is invaluable. Generating graphs for key counters is also mandatory for mature systems operation; one way to collect and display key counts for Brocade switches is to use Munin and a Munin plugin which I wrote.

PS: Brocade documentation states that SFP problems might result in the combination of rises in CRC errors and encoding/disparity errors. This did not happen in this situation.

Wednesday, June 23, 2010

Separation or co-location of database and application

In a classical three-tier architecture (database, application-server, client), a choice will always have to be made: Should the database and the application-server reside on separate servers, or co-located on a shared server?

Often, I see recommendations for separation, with vague claims about performance improvements. But I assert that the main argument for separation is bureaucratic (license-related), and that separation may well hurt performance. Sure, if you separate the application and the database on separate servers, you gain an easy scale-out effect, but you also end up with a less efficient communication path.

If a database and an application is running on the same operating system instance, you may use very efficient communication channels. With DB2, for example, you may use shared-memory based inter-process communication (IPC) when the application and the database are co-located. If they are on separate servers, TCP must be used. TCP is a nice protocol, offering reliable delivery, congestion control, etc, but it also entails several protocol layers, each contributing overhead.

But let's try to quantify the difference, focusing as closely on the query round-trips as possible.

I wrote a little Java program, LatencyTester, which I used to measure differences between a number of database<>application setups. Java was chosen because it's a compiled, statically typed language; that way, the test-program should have as little internal execution overhead as possible. This can be important: I have sometimes written database benchmark programs in Python (which is a much nicer programming experience), but as Python can be rather inefficient, I ended up benchmarking Python, instead of the database.

The program connects to a DB2 database in one of two ways:
  • If you specify a servername and a username, it will communicate using "DRDA" over TCP
  • If you leave out servername and username it will use a local, shared memory based channel.
After connecting to the database, the program issues 10000 queries which shouldn't result in I/Os, because no data in the database system is referenced. The timer starts after connection setup, just before the first query; it stops immediately after the last query.

The application issues statements like VALUES(...) where ... is a value set by the application. Note the lack of a SELECT and a FROM in the statement. When invoking the program, you must choose between short or long statements. If you choose short, statements like this will be issued:
VALUES(? + 1)
where ? is a randomly chosen host variable.
If you choose long,  statements like
will be issued, where lksjflkvjw...pvoiwepwvk is a 2000-character pseudo-randomly composed string.
In other words: The program may run in a mode where very short or very long units are sent back and forth between the application and the database. The short queries are effectively measuring latency, while the long queries may be viewed as measuring throughput.

I used the program to benchmark four different setups:
  • Application and database on the same server, using a local connection
  • Application and database on the same server, using a TCP connection
  • Application on a virtual server hosted by the same server as the database, using TCP
  • Application and database on different physical servers, but on the same local area network (LAN), using TCP
The results are displayed in the following graph:

Results from LatencyTester
Click on figure to enlarge; opens in new window

Clearly, the the highest number of queries per second is seen when the database and the application are co-located. This is especially true for the short queries: Here, more than four times as many queries may be executed per second when co-locating on the same server, compared to separating over a LAN. When using TCP on the same server, short-query round-trips run at around half the speed.

The results need to be put in perspective, though: While there are clear differences, the absolute numbers may not matter much in the Real World. The average query-time for short local queries were 0.1ms, compared to 0.8ms for short queries over the LAN. Let's assume that we are dealing with a web application where each page-view results in ten short queries. In this case, the round-trip overhead for a location connection is 10x0.1ms=1ms, whereas round-trip overhead for the LAN-connected setup is 10x0.8ms=8ms. Other factors (like query disk-I/O, and browser-to-server roundtrips) will most likely dominate, and the user will hardly notice a difference.

Even though queries over a LAN will normally not be noticeably slower, LANs may sometimes exhibit congestion. And all else being equal, the more servers and the more equipment being involved, the more things can go wrong.

Having established that application/database server-separation will not improve query performance (neither latency, nor throughput), what other factors are involved in the decision between co-location and separation?
  • Software licensing terms may make it cheaper to put the database on its own, dedicated hardware: The less CPUs beneath on the database system, the less licensing costs. The same goes for the application server: If it is priced per CPU, it may be very expensive to pay for CPUs which are primarily used for other parts of the solution.
  • Organizational aspects may dictate that the the DBA and the application server administrator each have their "own boxes". Or conversely, the organization may put an effort into operating as few servers as possible to keep administration work down.
  • The optimal operating system for the database may not be the optimal operating system for the application server.
  • The database server may need to run on hardware with special, high-performance storage system attachments. - While the application server (which probably doesn't perform much disk I/O) may be better off running in a virtual server, taking advantage of the flexible administration advantages of virtualization.
  • Buying one very powerful piece of server hardware is sometimes more expensive than buying two servers which add up to the same horsepower. But it may also be the other way around, especially if cooling, electricity, service agreements, and rack space is taken into account.
  • Handling authentication and group memberships may be easier when there is only one server. E.g., DB2 and PostgreSQL allows the operating system to handle authentication if an application connects locally, meaning that no authentication configuration needs to be set up in the application. (Don't you just hate it when passwords sneak into the version control system?)
  • A mis-behaving (e.g. memory leaking) application may disturb the database if the two are running on the same system. Operating systems generally provide mechanisms for resource-constraining processes, but that can often be tricky to setup.
Pro separationCon separation
Misbehaving application will not be able to disturb the database as much.Slightly higher latency.
May provide for more tailored installations (the database gets its perfect environment, and so does the application).Less predictable connections on shared networks.
If the application server is split into two servers in combination with a load balancing system, each application server may be patched individually without affecting the database server, and with little of no visible down-time for the users.More hardware/software/systems to maintain, monitor, backup, and document.
May save large amounts of money if the database and/or the application is CPU-licensed.Potentially more base software licensing fees (operating systems, supporting software).
Potentially cheaper to buy two modest servers than to buy one top-of-the-market server.Potentially more expensive to buy two servers if cooling and electricity is taken into account.
Allows for advanced scale-out/application-clustering scenarios.Prevents easy packaging of a complete database+application product, such as a turn-key solution in a single VMWare or Amazon EC2 image.
Am I overlooking something?
Is the situation markedly different with other DBMSes?

Monday, December 28, 2009

PostgreSQL innovation: Exclusion constraints

It seems like the next generation of PostgreSQL will have a new, rather innovative feature: Exclusion constraints. The feature is explained in a video from a presentation at a recent San Francisco PostgreSQL Users' Group; the presentation couples the new feature with the time period data type.

In a perfect World where databases implement the full ISO SQL standard (including non-core features), exclusion constraints could be nicely expressed as SQL assertions, but the perfect World hasn't happened yet. And I think that I know the reason for this: SQL assertions seem like a strong cocktail of NP-hard problems - they are probably very hard to implement in an efficient way.

In our less than perfect World, it's nice that PostgreSQL will soon offer a way to specify exclusion constraints other than the UNIQUE constraint.

The presenter, Jeff Davis, has an interesting blog, by the way. An on the subject of video, Vimeo has a little collection of PostgreSQL video clips that looks interesting.

Sunday, October 04, 2009

SQL comparison update: PostgreSQL 8.4

I finally got around to adjusting my SQL comparison page, so that the improvements in PostgreSQL 8.4 are taken into account. PostgreSQL is now standards-compliant in the sections about limiting result sets; actually, PostgreSQL is the first DBMS that I know of which supports SQL:2008's OFFSET + FETCH FIRST construct for pagination. PostgreSQL's new support for window functions is also very nice.

My page doesn't cover common table expressions (CTEs) yet, but it's certainly nice to see more and more DBMSes (including PostgreSQL, since version 8.4) supporting them. Even non-recursive CTEs are important, because they can really clean up SQL queries and make them more readable.

SQL comparison updates: Oracle 11R2, diagnostic logs

I finally found some time to update my page which compares SQL implementations. I performed a general update regarding Oracle, now that Oracle 11R2 has been released. And I added a new (incomplete) section which describes where the different DBMSes store diagnostic logs.

I turned out that very little has changed in Oracle since generation 10. The only remarkable new SQL feature in Oracle 11 is support for CTEs and proper CTE-based recursive SQL (introduced in version 11, release 2) -- but as I don't cover this topic on my page, updating the Oracle items was mostly a question of updating documentation links. Oracle still doesn't support the standard by having a CHARACTER_LENGTH and a SUBSTRING function, for example. This is simple, low-hanging fruit, Oracle(!) Sigh. It seems like Oracle's market position has made them (arrogantly) ignore the SQL standard.

Monday, August 31, 2009

VLDB2009: Non-cloud storage

Not all of VLDB2009 was related to cloud storage. Luckily, local and SAN storage is still being explored. Here are some notes from selected presentations.

Mustafa Canim: An Object Placement Advisor for DB2 Using Solid State Storage: What data should be put on disks, and what data on solid state disk storage? Solid state drives (SSDs) shine at random read I/O, but in most situations, there's limited funds, so only part of a database will be eligible for SSD placement. It turns out that if a naïve/simplified strategy like let's place all indexes on SSD is used, only a small overall performance gain is measurable, and it's hardly a justification of the expensive SSD storage. But if placement of database objects (tables/indexes) is based on measurements from a representative workload, then data which is often seeked to can be placed on the SSDs. Canim has created a prototype application which uses DB2 profiling information to give advice on the most optimal use for xGB of SSD storage, and he demonstrated very convincing price/performance gains from his tool. The principle should be easy to apply to other DBMSes.

Devesh Agrawal: Lazy-Adaptive Tree: An Optimized Index Structure for Flash Devices: Since SSDs do not shine at random writes, an SSD-optimized index structure would be highly welcome. The Lazy-Adaptive (LA) Tree index is a (heavily) modified B+ tree which buffers writes in a special way, yielding significant performance improvements.

Nedyalko Borisov demonstrating DIADSNedyalko Borisov and Shivnath Babu had a poster and a demonstration about a prototype application they have created: DIADS. DIADS integrates information from DB2 query access plans with knowledge about the storage infrastructure (all the way from files on the disk, through the LVM and SAN layers, to individual disks) to help diagnosing performance bottlenecks. This would certainly be useful for DBAs, and it could probably bridge the worlds of DBAs and storage people. They are considering making it an open source project. By the way: I believe that I've heard Hitachi claim to be selling a tool with a similar objective: Hitachi Tuning Manager.

Finally, a little rule of thumb from Patrick O'Neil's The Star Schema Benchmark and Augmented Fact Table Indexing (part of the TPC Workshop): Throughput of magnetic disks has grown much more than seek latency, so at the moment, 1MB of scanning can justify 1 seek.

Friday, August 28, 2009

VLDB2009: Sloppy by choice

One of the recurring themes at the VLDB2009 conference was how to create massively scalable database systems, by decreasing expressiveness, transaction coverage, data integrity guarantees—or any combination of these. The theoretical justification for this is partly explained by the CAP Theorem. Much has already been written about database key-value/object stores, cloud databases, etc. But there were still a few surprises.

Ramakrishnan's keynote
Yahoo!'s Raghu Ramakrishnan (whose textbook many readers of this article will know) gave a keynote on the subject. It was actually new to me that Yahoo! is also entering the cloud business; it's getting crowded in the sky, for sure. As we know, the business idea is this: A large operation like Amazon already has a massive, distributed IT infrastructure; so letting other companies in isn't that much of an extra burden. And the more users of the hardware, the easier it is to build a cost-efficient setup which can still handle spikes in performance demands (with many users it's very unlikely that their systems run at peak performance at the same time). Nice idea, but it may take a long while before users convert to using software which runs in the cloud, and it remains to be seen how many cloud vendors which can survive that long.

Anyway, Raghu Ramakrishnan presented a much-needed comparison of cloud database solutions (pages 55-60 in this presentation). The consistency model of Yahoo!'s cloud database system, PNUTS, does not provide ACID, but nor is it 'sloppy' to the degree of BASE. Another nice aspect of Yahoo!'s cloud systems is that much of it is based on open source software. Yay Yahoo!

When to be sloppy
At the conference, some claimed that cloud storage can also be used for important data, but no one gave a plausible example. In cloudy times, we should not forget that not all data are about tweets, status updates, weblog postings, etc. There are actually data which is actually important. That's why I liked the Consistency Rationing in the Cloud: Pay only when it matters presentation: It provided at framework for categorizing data in degrees of acceptable sloppiness, based on the cost associated with potential inconsistencies, versus the savings gained from the lower transaction overhead.

Panel on cloud storage
On Wednesday, there was a panel discussion on How Best to Build Web-Scale Data Managers, moderated by Philip Bernstein. Random notes from the discussion:
  • A Surprising, and somewhat strange viewpoint from Bernstein: We should not ditch ACID (not surprising coming from Mr Transaction himself), but we should give hierarchical DBMSes a new chance. According to Bernstein: The reason why it will not fail this time is that we have become so good at handling materialized views, and they allow us to make sure that fast queries are not restricted to restricting/scanning one one dimension. Bernstein failed to alleviate my fear of the return of another major drawback of hierarchical databases: navigational queries.
  • While there's much not-so-important data out there, the phenomenon of moderate amounts of important data hasn't gone away (not every business is a So although the non-ACID, non-relational database systems may have a lot of attention, it doesn't matter for the makers of "traditional" DBMSes, because RDBMS business is doing great.
  • Sub-question: Why do web start-ups seem to make use of key-value stores, and not use Oracle's DBMS (for example) when they need to scale to beyond a single data server? There wasn't much opposition to the view that—in addition to being administration labor intensive—the cost of an Oracle cluster is way out of budget in many businesses. So: If database researchers want to help prevent relational+transactional research from becoming increasingly irrelevant, it's time to help the open source database projects. While I agree that researchers can make a difference in the open source world, but I's skeptical to the perception of RDBMSes being abandoned; whatever numbers I've seen actually indicate the opposite. And while Facebook—for example—has a key-value store, their non-clickstream-data is actually in sharded MySQL databases, as far as I've heard; sharded MySQLs will never win relational database beauty contests, but at least it's tabular data, accessible with SQL, and with node-local transactions.
  • Interesting point: SAP, an application with undisputed business significance, is well known for using nothing but the most basic RDBMS features. With that in mind, one should be cautious to denounce cloud databases for lack of expressiveness.

Finally, a couple of pictures from the nearby Tête d'Or Park:

VLDB2009: TPC Workshop

Monday, I attended the Transaction Processing Council (TPC) Workshop about the latest developments in database benchmarks. The workshop was kicked off with a keynote from a true database hero, Michael Stonebraker. Mr. Stonebraker has not only published significant research papers—he is also initiated a number of projects and startups: Postgres (the precursor to the great PostgreSQL DBMS), Vertica (one of the pioneers within the "column-oriented"/"column-store" database realm), StreamBase, and others. Stonebraker held it that benchmarks have been instrumental in increasing competition among vendors, but there are aspects to be aware of: As the benchmarks allow the vendors to tune the DBMS (and sometimes create setups not resembling most setups, like using five figure disk counts), this doesn't improve the out-of-box experience—an experience which all too often needs to be significantly tweaked. Stonebraker also criticized the TPC from being too vendor focused, instead of having focus on users and (simple) applications. And he urged the TPC to speed up the development of new benchmark types (I'm thinking: geospatial, recovery, ...), partly by cutting down on organizational politics.

Personally, I'm astonished that some (most?) of the big DBMS vendors prohibit their users from publishing performance findings. This curbs discussion among practitioners, and it decreases reproducibility and detail of research papers ("product A performed like this, product B performed like that"). I doubt that this actually holds in a court of law, but it would certainly take guts (and a big wallet) to challenge it. I'm also annoyed that the vendors don't really support the TPC much: The TPC-E benchmark (OLTP-benchmark, sort-of modernized version of TPC-C) is two years old, yet only one vendor (Microsoft) has published any results yet.

Nevertheless, references to TPC benchmarks were prevalent at the conference, being referred to in several papers.

I'm planning to try running TPC-H on our most important database systems, to see if it is feasible to use it for regular performance measurements—in order to become aware of performance pro-/re-gressions. By the way: It would be of great if IBM (and others) published a table of reference DB2 TPC findings for a variety of normal hardware configurations. That way, a DBA like me could get an indication of whether I've set up our systems properly.

Other speakers had various ideas for new benchmarks, including benchmarks which measure performance per kWh, and benchmarks which expose how well databases handle error situations.

A researcher from VMWare pledged for benchmarks of databases running in virtual environments. He presented some numbers of a TPC-C-like workload running on ESX guests, showing that a DBMS (MSSQL, I believe) can be set up to run at 85%-92% of the native speed. Certainly acceptable. And much in like with what I'm experiencing. I hope that figures like this can gradually kill the myth that DBMSes shouldn't run in virtual hosts—a myth which results in a situation where many organizations don't realize the full potentials of virtualization (increased overall hardware utilization/lower need for manpower/less late-night service windows, as workloads can be switched to other hosts when a server needs hardware service).

I forgot to take a picture from the workshop, so here's a picture of me being sceptical about traditional DBMS vendors—next to the Saôme river.

At VLDB2009

I'm attending the 35th VLDB conference in Lyon: VLDB2009. The conference portrays itself as the premier international forum for database researchers, vendors, practitioners, application developers, and users. Of the 700 people (from 44 countries), I'm one of the few practitioners at the conference; and though there's a risk that the conference will be too research oriented, I've signed up, especially hoping to get the latest updates and thoughts on
  • probabalistic databases
  • column-oriented databases
  • performance quantification
  • cloud databases
During the next couple of days, I'll share my experiences here.

I—and others—have tweeted a bit from the conference, as well.